SFRP2 induces a mesenchymal subtype transition by suppression of SOX2 in glioblastoma

Intratumoral heterogeneity is a characteristic of glioblastomas that contain an intermixture of cell populations displaying different glioblastoma subtype gene expression signatures. Proportions of these populations change during tumor evolution, but the occurrence and regulation of glioblastoma subtype transition is not well described. To identify regulators of glioblastoma subtypes we utilized a combination of in vitro experiments and in silico analyses, using experimentally generated as well as publicly available data. Through this combined approach SOX2 was identified to confer a proneural glioblastoma subtype gene expression signature. SFRP2 was subsequently identified as a SOX2-antagonist, able to induce a mesenchymal glioblastoma subtype signature. A subset of patient glioblastoma samples with high SFRP2 and low SOX2 expression was particularly enriched with mesenchymal subtype samples. Phenotypically, SFRP2 decreased tumor sphere formation, stemness as assessed by limiting dilution assay, and overall cell proliferation but increased cell motility, whereas SOX2 induced the opposite effects. Furthermore, an SFRP2/non-canonical-WNT/KLF4/PDGFR/phospho-AKT/SOX2 signaling axis was found to be involved in the mesenchymal transition. Analysis of human tumor tissue spatial gene expression patterns showed distinct expression of SFRP2- and SOX2-correlated genes in vascular and cellular areas, respectively. Finally, conditioned media from SFRP2 overexpressing cells increased CD206 on macrophages. Together, these findings present SFRP2 as a SOX2-antagonist with the capacity to induce a mesenchymal subtype transition in glioma cells located in vascular tumor areas, highlighting its role in glioblastoma tumor evolution and intratumoral heterogeneity.


Introduction
Glioblastoma is the most aggressive brain tumor type in adults and lacks effective therapeutic options. Glioblastomas are divided into three gene-expression signature-based subtypes denoted proneural, mesenchymal, and classical [1,2]. Single-cell analyses have revealed that individual glioblastomas contain intermixtures of cells displaying these signatures, suggesting a logical connection between observed intertumoral and intratumoral heterogeneity [3].
The intratumoral glioblastoma heterogeneity is well reported in multiple instances [1, [4][5][6][7] and genetically different cell populations have been shown to co-evolve and to be inter-clonally dependent [8]. This is considered to be a major hurdle for successful treatment of glioblastoma, where cells may display differential drug sensitivities, thereby posing an increased risk for developing treatment resistance and the pre-existence of resistant subpopulations, as exemplified in attempts of anti-EGFR treatment and in temozolomide resistance [9][10][11]. Intratumoral heterogeneity among the malignant cells may arise due to variation of lines and levels of differentiation, coevolution of populations with different genetic or epigenetic characteristics, local microenvironmental cues or a combination of these that may evoke different phenotypes [12][13][14]. Frequently occurring genetic alterations in glioblastoma have been connected with different phenotypic states, and cells may also appear in intermediate hybrid states due to plasticity within single tumors [15]. Furthermore, a proneural to mesenchymal transition has been associated with a worse patient outcome and an attraction of microglia cells [2,3,16]. The transitions between glioblastoma gene expression subtypes and the spatial organization of different subtype cells in the tumor therefore needs to be further studied for fundamental biological understanding and development of effective treatment against this devastating disease.
We have previously established and classified highgrade glioma cell cultures into two subtypes based on their gene expression patterns, denoted type A and type B cultures [17]. Type A cultures form tumor spheres, grow as intracranial xenografts and express astroglial and stem cell-related genes including GFAP (Glial Fibrillary Acidic Protein), SOX2 (SRY (sex determining region Y)-box 2) and PDGFRA (Platelet Derived Growth Factor Receptor α). Type B cultures are defined by a low capacity to grow as tumor spheres and to generate intracranial xenografts, and express mesenchymal-related genes including FN1 (Fibronectin), CXCL12, and PDGFRB (Platelet Derived Growth Factor Receptor β). In clinical material, genes highly expressed in type A cultures were correlated with the non-mesenchymal glioblastoma subtype signatures, proneural and classical, whereas genes expressed in type B cultures correlated with the mesenchymal subtype [17]. Since an intermixture of glioblastoma signature displaying cells have been described to occur within individual tumors [3], we herein sought to identify regulators of these signatures and their spatial organization in glioblastomas. To achieve this we used a combined approach of in vitro experiments including overexpression screens, phenotypic profiling, co-immunoprecipitation with subsequent mass spectrometry, and RNA-sequencing, as well as in silico analyses of publicly available data including connectivity map (CMap) and spatial gene expression analyses.

Results
Identification of SOX2 and SFRP2 as potential glioblastoma subtype regulators To experimentally identify potential regulators of glioblastoma gene expression subtype signature we performed two screens by crosswise overexpressing type A genes in a type B culture or type B genes in a type A culture (Fig. 1A). These screens were based on the hypothesis that certain genes in a signature are the ones that have the actual causal effect on the signature. We selected the Hesselager cultures U-2987 (cell culture 18) and U-2982 (cell culture 11) as representative type A and B cultures, respectively (Fig. 1A, B) [17].
To illustrate what glioblastoma subtype these cultures represent their gene expression signatures were compared to other glioblastoma data sets. As previously shown, the type A culture U-2987 expressed high SOX2 and low FN1, whereas the type B culture U-2982 displayed an opposite expression pattern (Fig. 1B) [17]. The type A and B gene expression patterns, arranged by the correlation to FN1-expression (Fig. 1B), were also observed among Cancer cell line encyclopedia (CCLE) glioma cell lines (Fig. S1A) and The cancer genome atlas (TCGA) glioblastoma samples (Fig. S1B) [18]. Finally, the U-2987 and U-2982 expression signatures were compared with the recently updated glioblastoma subtype gene expression signatures [2]. Highly expressed genes in U-2987 were enriched in proneural subtype signature genes, 14 out of 50, and in contrast genes highly expressed in U-2982 were enriched in mesenchymal subtype signature genes, 25 out of 50 (Fig. 1C). Thus U-2987 and U-2982 are cultures representative of proneural and mesenchymal glioblastoma subtype signatures, respectively, and will be referred to as such throughout this work.
In our first screen of a mesenchymal-to-proneural transition, type A genes (Table S1) were overexpressed in the mesenchymal culture U-2982, where decreased FN1-expression was used as a proxy indicator for a mesenchymal-to-proneural signature change. Here, SOX2 strongly suppressed FN1, and was thus considered as a potential mesenchymal-to-proneural transitioninducing gene (Fig. 1D).
Since glioblastoma tumor progression is suggested to proceed from a proneural-like to a mesenchymal subtype [19], our second screen was aimed to identify genes with a capacity to induce a proneural-to-mesenchymal subtype transition. Based on the findings in our first screen, and the low expression of SOX2 in mesenchymal subtype we sought to identify genes that can suppress SOX2 in U-2987. Secreted frizzled-related protein 2 (SFRP2), Krüppel-like factor 4 (KLF4), and Regulator of G protein signaling 5 (RGS5) were found to strongly decrease SOX2 expression, out of 17 screened type B genes ( Fig. 1E and Table S1). SFRP2 overexpression and decreased SOX2 levels were respectively confirmed and recapitulated by qPCR analysis in newly transduced cells (Fig. 1F). Since SFRP2 showed the highest antagonistic effect on SOX2 expression and had a high novelty value it was selected for further experimentation.
In parallel to the in vitro screens an in silico CMap analysis was performed to identify potential regulators of gene expression signatures. To decrease bias due to cell culture conditions, we extracted gene expression signatures from cell line panels that had been cultured in serum or in neurosphere conditions, CCLE [18] and the Human glioblastoma cell culture resource (HGCC) [20], respectively. Using unsupervised analysis CCLE and HGCC glioma cell lines were hierarchically clustered based on their most differentially expressed genes, yielding two main branches per dataset, denominated C1 and C2, and H1 and H2, respectively (Fig. 1G, Table S2). Subsequently, the 100 highest positively and negatively differentially expressed genes between C2 and C1, or H2 and H1, were used as master signatures for queries in the CLUE (CMap and LINCS Unified Environment) database. A SOX2 CMap signature was identified to have the highest signature similarity score for both master signatures (Fig. 1G). That is, the top prediction in CLUE, for the signature genes distinguishing the C2 and C1, and the H2 and H1 branches, was that these were contributed to by SOX2 regulation. In support of this, average SOX2 expression was significantly higher in C2 as compared to C1 (p = 0.02), and in H2 versus H1 (p = 0.04) (Fig. 1G). Furthermore, the H1 branch was enriched with cell lines of mesenchymal glioblastoma subtype (22/27), and the H2 branch with proneural and classical subtypes (15/21). Together these in vitro and in silico analyses suggested SOX2 as a top regulator of glioma cell culture gene expression signatures irrespective of growth conditions. SFRP2 induces a proneural to mesenchymal glioblastoma subtype transition Upon further analysis of TCGA glioblastomas SFRP2 was found to be significantly higher expressed in mesenchymal as compared to proneural and classical glioblastoma subtypes ( Fig. 2A). To investigate if SFRP2 may induce a mesenchymal subtype signature, the proneural cell culture U-2987 was subjected to RNA-sequencing analysis upon SFRP2overexpression (Fig. 2B). Indeed, SFRP2-overexpression shifted the subtype support index of this culture from a proneural to a mesenchymal signature (Fig. 2C). In comparison with CCLE and HGCC cluster master signature genes, SFRP2 increased the average expression of C1-and H1-genes by 0.68 and 0.30, and decreased C2-and H2-genes by −0.21 and −0.52, respectively (Fig. 2D). In vitro phenotype analyses after SFRP2-overexpression in U-2987 revealed significantly decreased tumor sphere formation, self-renewal capacity, as well as cell proliferation, and increased Matrigel invasive capacity ( Fig. 2E-H).
In the analyses of cell cycle regulator genes upon SFRP2 and SOX2-overexpression in U-2987 and U-2982, respectively, we found an opposite regulation of multiple cell cycle regulator genes, including CCNE1, at both transcript and protein level, and CDK2 at a transcript level ( Fig. 2I and Fig. S2A-C). SFRP2 decreased and increased the ratio of cells in G1-and S-phase, respectively ( Fig.  S2D), consistent with a de-regulated G1/S cell cycle control. Conversely, SOX2 increased and decreased the ratio of cells in G1-and S-phase, respectively (Fig. S2E). SOX2 has mechanistically been connected to cell cycle regulation by becoming stabilized by G1 cyclins (CCND1 and CCNE1) and their associated cyclin-dependent kinases, including CDK2 [21]. Here we found that SFRP2overexpression in U-2987 rendered the cells less sensitive to CDK1/2 inhibition by CVT-313 (Fig. 2J). Consistent with a destabilizing effect on SOX2, possibly related to a decrease in G1-S cyclin/CDK proteins, CVT-313 decreased SOX2 protein level in U-2987, whereas SOX2 protein level was already decreased in U-2987 SFRP2overexpression cells (Fig. 2K). In parallel, SFRP2overexpression in U-2987 increased the apoptotic response to Staurosporin, as assessed by level of cleaved PARP (Fig. 2L). Fig. 1 Identification of SOX2 and SFRP2 as antagonistic regulators of gene expression differences between glioblastoma cell lines. A Cartoon outlining the setup of the overexpression screens. The arrows indicate the screening of type B genes to identify those that can transition a type A culture into a type B signature, and conversely screening of type A genes to identify those that can transition a type B culture into a type A signature. In this screen, FN1 was used as a proxy for a type B signature and SOX2 as a proxy for a type A signature. B Heatmaps illustrating gene expression patterns of type A and type B genes in Hesselager cultures. Cell cultures are arranged from low-tohigh expression of FN1, followed by a high-to-low arrangement according to GFAP expression within the high or low FN1 expressing groups. In the heatmap, the genes were arranged from left-to-right in descending order according to their overall correlation to FN1. The red and blue colors indicate relatively high and low expression levels, respectively. The black and white boxes on top of the heatmap depict type A and type B genes, respectively. C Venn diagrams illustrating results from overlap analyses between relatively high (>2-fold) expressed genes in U-2987 versus U-2982 to the left and U-2982 versus U-2987 to the right, and glioblastoma subtype expression signature genes. Number in brackets below each cell line name indicates number of relatively high expressed genes per corresponding cell line. Gene lists indicate overlapping genes for the corresponding highest overlap. D Overexpression screen results where FN1 gene expression levels were monitored by qPCR upon overexpression of mentioned type A genes (Table S1) in culture U-2982. E Overexpression screen results where SOX2 gene expression levels were monitored by qPCR upon overexpression of mentioned type B genes (Table S1) in culture U-2987. The numbering after certain genes, refer to the clone number of used ORFs when several ORFs were cloned. F Expression of SFRP2 and SOX2 analyzed by qPCR upon SFRP2-overexpression in U-2987. G The left and right heatmaps with corresponding dendrograms depict results of hierarchical clustering analysis based on differentially expressed genes of cell lines from CCLE glioma and HGCC, respectively. The highest order cluster branches of the CCLE glioma and HGCC cell lines are denominated C1 and C2, and H1 and H2, respectively. The table in the middle panel lists the top scoring genes from CMap analyses based on differentially expressed genes between cluster branches C1 and C2, and H1 and H2, respectively. The bar graph below each heatmap shows the SOX2 expression level in each corresponding cell line (z-score). For the HGCC cell lines the corresponding predicted glioblastoma gene expression subtype is denoted as reported in the HGCC database. For western blot analyses GAPDH levels were added to assess protein input levels in indicated panels.
In connection to previously reported glioblastoma mesenchymal signature transcription factors, SFRP2 increased and correlated with the expression of CEBPB, RUNX1, and FOSL2 in TCGA glioblastomas (Fig. S3A, B) [22].
Collectively, SFRP2 high expressing glioblastomas were enriched with mesenchymal subtype tumors, and in vitro SFRP2 induced a proneural-to-mesenchymal subtype gene expression signature transition accompanied by decreased stem cell features, de-regulated proliferation, as well as increased Matrigel invasiveness and sensitivity to apoptotic stimuli. This indicated SFRP2 as an inducer of mesenchymal glioblastoma subtype.
Regarding previously reported glioblastoma mesenchymal signature transcription factors, SOX2 as opposed to SFRP2, decreased CEBPB and FOSL2 levels consistent with their negative correlation to SOX2 in TCGA glioblastomas (Fig. S3A, B). In tumor subgroups based on SOX2 and SFRP2 transcript levels in the TCGA and The Chinese gliomas genome atlas (CGGA) glioblastoma datasets [23], the respective SFRP2 high /SOX2 low subgroups were enriched with mesenchymal subtype tumors (Fig. 3H, I).
In summary, SOX2-overexpression mirrored the SFRP2overexpression phenotype and a subgroup of glioblastomas with high SFRP2 and low SOX2 expression was enriched with mesenchymal subtype tumors. SOX2 levels are controlled by an SFRP2/noncanonical-WNT/KLF4/PDGFR/phospho-AKT signaling axis SFRP2 is a secreted protein reported to prevent WNTligands from binding to Frizzled receptors and thereby modulate WNT-signaling [24]. Upon investigation of canonical WNT signaling, SFRP2-overexpression indeed decreased level of active β-catenin (Fig. 4A). However, shRNA suppression of β-catenin did not affect SOX2 levels per se. SFRP2 may also act via non-canonical WNT-signaling, including the ROR2 receptor (receptor tyrosine kinase like orphan receptor 2) [25]. No ROR2-SFRP2 interaction could be identified by co-immunoprecipitation (Fig. 4B), but ROR2 expression actually increased upon SFRP2-overexpression in U-2987 (Fig. 4C, D), and subsequent suppression of ROR2 could not restore SOX2 expression to any noticeable degree (Fig. 4E). Together, this did not support a direct involvement of β-catenin or ROR2 in the SFRP2 effect on SOX2.
To identify novel intermediate regulatory factors between SFRP2 and SOX2 we hypothesized that activity of such genes should increase or decrease upon SFRP2overexpression but not upon SOX2-overexpression, and correlate with SOX2 expression in the CCLE glioma cell lines. We found that the transcription factors POU3F2 (OCT7) and KLF4 fulfilled these criteria (Fig. 4F). POU3F2 decreased upon SFRP2-overexpression and correlated positively to SOX2 expression in the CCLE glioma cell lines. KLF4 displayed an opposite pattern where it increased upon SFRP2-overexpression and correlated negatively to SOX2 expression (Figs. 4F and S4). Experimentally, KLF4 indeed increased upon SFRP2overexpression (Fig. 4G). Direct KLF4-overexpression also decreased SOX2 levels, supporting its involvement (Fig. 4G), and similar to SFRP2, KLF4-overexpression increased FN1 protein levels, suggesting a mesenchymal transition. Conversely, suppression of POU3F2 did not have any effect on SOX2 levels. POU3F2 instead appeared subordinate to both SFRP2 and KLF4, just as SOX2, since both SFRP2-and KLF4-overexpression decreased POU3F2 and SOX2 protein levels (Fig. 4G). In other instances, KLF4 and SOX2 have been connected to PI3K signaling [26]. Consequently, addition of the AKTinhibitor MK2206, decreasing the phosphorylation of AKT at serine residue 473, resulted in decreased SOX2 levels (Fig. 4H). Finally, both SFRP2-and KLF4overexpression decreased AKT phosphorylation at serine residue 473 (Fig. 4H). Together, this suggests that the SOX2 protein level is maintained by phosphorylated AKT, and that AKT phosphorylation can be decreased by SFRP2 and KLF4, leading to decreased SOX2 levels. This is in agreement with that SOX2 expression is connected to AKT signaling [27], and that KLF4overexpression decreases AKT phosphorylation [28].
Upon SOX2-overexpression in culture U-2982 and subsequent phosphoproteomic PamGene analysis, the phosphorylation pattern shifted to resemble the one found in U-2987 cells (Fig. 4I, J, Table S3). The mesenchymalto-proneural glioblastoma subtype signature transition is thus paralleled by a tyrosine phosphorylation-signature shift. The SOX2-induced phosphorylation pattern was predicted to stem from PDGFR-alpha receptor signaling (Fig. 4J). Indeed, PDGFRA protein appeared upon SOX2-overexpression in U-2982 (Fig. 4K), while SFRP2 in U-2987, in an opposing manner but to a lower extent, decreased PDGFRA levels (Fig. 4K). In addition, PDGFRA transcript levels also decreased, and were tracked by decreased PDGFB ligand transcript levels (Fig. 4L). Conversely, transcript levels of PDGFA, PDGFC and PDGFRB increased, suggesting a shifted PDGFR-signaling mode. Finally, addition of the PDGFRinhibitor CP-673451 to U-2987 decreased SOX2 protein levels (Fig. 4M). In summary, SOX2-overexpression increased PDGFRA protein levels and induced a PDGFRA phosphorylation signature, while use of inhibitors of either PDGFR or AKT decreased SOX2 protein levels. On the other hand, SFRP2-overexpression increased KLF4 expression and decreased PDGFRA transcript and protein levels, and both SFRP2-and KLF4overexpression decreased AKT phosphorylation at serine residue 473.
To identify SFRP2 interaction partners in U-2987 with SFRP2-overexpression we utilized an SFRP2 coimmunoprecipitation and subsequent mass spectrometry analysis approach (Table S4). In the SFRP2 immunoprecipitate  Fig. 2A but for SOX2 expression. B-F Same as in Fig. 2B, C, E, F, H, but for SOX2-overexpression in the type B culture U-2982. G Western blot experiment to illustrate FN1 and SFRP2 levels in U-2987 with SFRP2-overexpression, with and without re-introduced SOX2. H Bar graphs showing the 142 TCGA glioblastoma samples divided into four groups based on their relative gene expression of SFRP2 and SOX2, where the cutoffs are based on mean expression (black bars for SFRP2 with median z-score = −0.198, and grey for SOX2 with median z-score = 0.076). The heatmap below shows the corresponding described gene expression subtype for each sample in the bar graph. The subtypes are color coded in the heatmap as indicated. The yellow lines separate groups as depicted. The Line plot in the right describes the percent distribution of glioblastoma subtypes within the SFRP2-and SOX2-based subgroups. I Same as Fig. 3H, but for the 180 glioblastoma samples from CGGA database (black bars for SFRP2 with median z-score = 1.3, and grey for SOX2 with median z-score = 7.01).
we identified SFRP2 itself and 622 further proteins (Fig. 4N and Table S4). Using an ontology enrichment analysis 88 significant ontologies were identified (Table S5). Three of these ontologies were related to non-canonical WNTsignaling ( Fig. 4O) with several proteins being in common (Fig. 4O).
In summary, we conclude that the SFRP2-induced SOX2 decrease and mesenchymal transition, occur through a signaling path including non-canonical WNT signaling, KLF4, PDGFRs, and phospho-AKT, comprising both signal transduction events and transcriptional regulation (Fig. 4P).

SFRP2-expression distributes in vascular areas of glioblastoma tissue
To investigate the spatial distribution of SFRP2 and SOX2 cell populations in an intratumoral heterogeneity perspective we analyzed publicly available gene expression data from laser capture-dissected areas in glioblastomas [29] and compared these patterns with the SFRP2-and SOX2regulated genes identified in our in vitro experiments (Table S6). SFRP2 was significantly higher expressed in HBV (Hyperplastic blood vessels) and MVP (Microvascular proliferation) areas versus CT (Cellular tumor) areas (Fig. 5A). Furthermore, the top 50 positively SFRP2-correlated genes in the tissue were highly expressed in HBV and MVP areas (Fig. 5B). Conversely, SOX2 displayed an opposite pattern (Fig. 5C). The top 50 SOX2 positively correlated genes were highly expressed in CT area (Fig. 5D). In a focused analysis of genes differentially expressed between VA (Vascular Areas: composed of HBV and MVP) and CT areas (Fig. 5E, Table S7) we found that 238 out of 619 (38%) VA genes were increased by SFRP2, out of which 84 also overlapped with SOX2-decreased genes from our experiments (Fig. 5F). Conversely, 119 out of 301 (40%) CT genes were decreased by SFRP2, and 139 (46%) increased by SOX2, with an overlap of 76 genes (Fig. 5G).
In conclusion, genes highly expressed in VA areas overlap with SFRP2 increased genes, whereas genes highly expressed in CT areas overlap with SOX2 increased genes.
By analyzing glioma tissue stainings from the Human Protein Atlas, a set of type A genes were found expressed in CT areas, whereas type B genes were found expressed in vessel related structures (Fig. 5H). In summary, these tissue analyses indicate that SFRP2 and its correlated genes and regulation in tissue and in vitro show both RNA and protein distribution in VA areas, whereas RNA and proteins for genes related to SOX2 are distributed in CT areas. This implies that mesenchymal subtype glioblastoma cells preferentially reside in vascular tumor areas and proneural subtype cells in cellular tumor areas.
Experimentally, conditioned media from SFRP2overexpressing cells increased CD206 (MRC1) expression on macrophages, whereas SOX2-overexpression conversely diminished CD206 expression (Fig. 6C). Gene expression analysis supported this increase of MRC1, whereas other macrophage markers including ARG, CD80, IL-10 and IL-12 showed minor changes (Fig. 6D). Tissue data from the Human Protein Atlas showed a confinement of macrophage proteins in The blue and red circles indicate number of detected proteins, where the blue circle area denotes the number of proteins detected in the control precipitation, the overlap between the circles the number of proteins detected in both precipitates and the red circle area the number of proteins only detected by the SFRP2 co-immunoprecipitation. The list shows the top 50 proteins according to area score and with more than 4 unique peptide hits. O The bar diagram shows selected ontology hits for WNT-related signaling from a query of the proteins listed in panel (N). P Summary of proposed mechanism of the SFRP2/noncanonical WNT/KLF4/PDGFR/phospho-AKT/SOX2 axis during the mesenchymal transition induced by SFRP2. In all western blots, GAPDH was used to assess protein input levels. vascular areas, consistent with the distribution of SFRP2related gene expression in glioblastoma tissue described above (Fig. 6E). Together this suggests an involvement of SFRP2 in the effect of mesenchymal subtype cells on pericytes and macrophages in glioblastoma.

Discussion
This work presents SFRP2 and SOX2 as counteracting inducers of gene expression based glioblastoma subtypes, where SFRP2 suppresses SOX2 via non-canonical WNT Venn diagram (middle) shows the number of overlapping genes between SFRP2-increased and SOX2-decreased genes. G Strip charts showing the number of genes altered by SFRP2-(left) and SOX2-(right) overexpression among the genes higher expressed in CT than VA area (>2-fold). Venn diagram (middle) shows the number of overlapping genes between SFRP2-decreased and SOX2-increased genes. H Immunohistochemical staining patterns of selected type A genes in CT areas and type B genes in VA areas in glioma tumor samples from the Human Protein Atlas. signaling, KLF4, PDGFRs, and AKT. SFRP2 transitions cells from a proneural into a mesenchymal glioblastoma gene expression signature, suggesting its involvement in glioblastoma tumor cell plasticity and tumor progression.
In this study SOX2 was initially identified as an inducer of glioblastoma subtype transition in an experimental overexpression-screen, and subsequently SOX2 was found through a CMap analysis to be at the apex of gene expression determination in glioblastoma cultures [31,32]. SOX2 is a well-described stem cell transcription factor that together with KLF4, OCT4 and MYC constitute the Yamanaka factor quartet that together can induce pluripotent stem cells [33]. SOX2 is expressed in neurogenic regions in human brain including the subventricular zone and in glioblastoma SOX2 is essential for maintaining a tumor-and sphere forming cell phenotype [17,34]. It is also in combination with POU3F2, SALL2, and OLIG2, able to induce glioblastoma tumor growth [35]. Interestingly, we find in this study that both SOX2 and POU3F2 appear subordinate to both SFRP2 and KLF4. SOX2 protein can be stabilized via phosphorylation by a CDK2/CCNE1 complex [21]. We here show that SFRP2-overexpression decreases CCNE1, CCND1 and CCNB1, and at the same time decreases SOX2 protein and overall cell proliferation in glioblastoma cells. We suggest that inactivation of the CDK-inhibitor p27(KIP) may be important here. P27 binds to and prevents activation of CDK2/CCNE1 or CDK4/ CCND1 and is involved in blocking the cell cycle in the G0/ G1 phase [36]. Additionally, p27 can be inactivated via AKT activation [37].
SFRP2 is a secreted factor reported to antagonize WNT signaling by sequestering of WNT ligands and has been connected with increased tumor growth, metastasis and therapy resistance [38], but has also been described as a tumor suppressor [24]. In this study we found that SFRP2 suppressed β-catenin, consistent with recent findings [39]. However, the decreased SOX2 levels observed here appeared independent of active β-catenin. Nor did we find any experimental support of SFRP2 acting via ROR2 [25] to regulate SOX2 levels. Although, we noted that ROR2 transcript and protein levels actually increased upon SFRP2-overexpression. In our co-immunoprecipitation and mass spectrometry analysis for SFRP2-interactors we instead detected other non-canonical WNT-signaling components including VANGL1, a transmembrane protein associated with non-canonical WNT-signaling recently reported to alter cell motility in glioblastoma [40].
Through a combined analysis approach of gene expression pattern and gene regulation data, KLF4 was identified as an intermediate factor between SFRP2 and SOX2, increased by SFRP2, subsequently leading to decreased SOX2 gene expression and protein levels. As a Yamanaka factor [33] KLF4 has been reported to associate with SOX2 to regulate gene transcription and induce pluripotency [41]. However, in glioblastoma KLF4 expression appears negatively correlated with SOX2 expression. KLF4 is higher expressed in mesenchymal, type B, glioblastoma cultures and was here found to act as a SOX2 antagonist.
Based on the data on PDGF ligands and receptors in our RNA-sequencing and PamGene data, confirmed by western blot analyses, the PDGFR/PI3K/AKT signaling axis caught our attention. With regard to previous observations of an active PDGFRA signaling loop in glioblastoma [42] we here experimentally show this can be maintained by SOX2, and subsequently turned off by SFRP2, potentially acting via KLF4. Furthermore, phosphorylation of AKT at serine 473 has been reported to regulate SOX2 protein levels and self-renewal activity [43,44]. Here AKT phosphorylation decreased upon SFRP2-or KLF4-overexpression, and treatment by AKT inhibitor MK2206 or PDGF inhibitor CP-673451 both decreased SOX2 protein levels. This adds the inhibition of the PDGFR-AKT pathway as key for the antagonistic effect of SFRP2 on SOX2. Furthermore, we find it notable that PDGFRA and GFAP proteins were highly expressed in the proneural culture U-2987 and decreased upon the SFRP2 induced mesenchymal transition. Pdgfra + /Gfap + neural stem cells in the subventricular zone of the adult mouse brain have been identified as the origin of PDGF-induced gliomas [45]. The proneural culture U-2987 also expressed high levels of PTPRZ1, which together with PDGFRA is a marker of oligodendrocyte precursors [46], indicating a mixture of astrocyte and oligodendrocyte progenitor characteristics.
Our study also highlights the significance of spatial tumor cell distribution. The analysis of SFRP2-and SOX2regulated and correlated genes revealed their distinct allocation in vascular or cellular areas, respectively, providing a clue to how intratumoral heterogeneity may evolve. Our experimental findings of SFRP2 and SOX2 signaling events represent a scenario whereby SFRP2 counteracts SOX2 function allowing glioma cell progression from a proneural towards a mesenchymal phenotype. In concert with our observed SFRP2-induced pericyte profile, SOX2 expressing glioblastoma cells have been shown in xenograft models to be able to lose SOX2 expression to generate tumor vessel cell structures that resemble pericytes [47]. Furthermore, we found higher levels of macrophage, pericyte, and proangiogenic factors in SFRP2 high areas, consistent with the vascular distribution of expressed SFRP2 and correlating genes, suggesting an effect of SFRP2 on the surrounding microenvironment. Other secreted factors from SFRP2overexpressing cells, including CXCL12 have been shown important for the recruitment of immune cells in glioblastoma [48]. However, what triggers the increased SFRP2 levels remains to be identified.
In conclusion, this work suggests an important role of SFRP2 in the transition from a proneural to a mesenchymal glioblastoma subtype via suppression of SOX2 during tumor progression. A role in intratumoral heterogeneity is suggested by the spatial distribution or SOX2 and SFRP2 expressing cells in glioblastoma tumors. In extension, we suggest that combined targeting of SFRP2 and SOX2 signaling should be investigated to achieve combined anti-angiogenic, anti-immunogenic and antiproliferative treatment effects.

Cell culture
The Hesselager glioma cultures 11 and 18 have respectively been referred to as U-2982 and U-2987 in other studies [16,49] and are here referred to as such for coherence. Cells were maintained in Minimum Essential Medium (MEM, Gibco) containing 10% FBS and 1% penicillinstreptomycin at 37°C with 5% CO 2 , except in sphere formation assay and limiting dilution assay, where cells were cultured in neurosphere medium (#05751 Neurocult NA-S proliferation human kit; StemCell Technologies), supplemented with 20 ng/ml EGF (Invitrogen), 20 ng/ml bFGF (Invitrogen), and 2 μg/ml heparin (Sigma). The identities of the cell cultures were confirmed by STR profiling at NGI-Uppsala, SciLifeLab, Uppsala University, using the AmpFISTR Identifiler PCR Amplification kit (Thermo Fisher) (Table S8). For further information please see supplemental material and methods.

Generation of stable cell cultures overexpressing either type A or type B genes
Fifteen type A genes and thirty type B genes were selected based on their differential expression between 23 highgrade glioma cultures according to methods previously described [50] (Table S1). For further information about overexpression screens, establishment of gene overexpressing or knockdown cell cultures, please see supplemental material and methods.
RNA sequencing, qPCR, and Western blot RNA-sequencing was performed as previously described [16] and data is deposited at EBI ArrayExpress (E-MTAB-7591). Detailed information on qPCR and Westernblot is found in supplemental material and methods. Primers and antibodies are shown in Table S10.
Glioblastoma gene expression subtype signature support index SFRP2-or SOX2-regulated genes were compared with subtype-defining centroid genes as previously described [1,16]. For further information see supplemental material and methods.

Phenotype assays
Cell phenotype assays, including tumor sphere formation, limiting dilution assay, Matrigel invasion, cell proliferation assay, and cell cycle analysis were performed. For detailed information please see supplemental material and methods.

Global protein tyrosine kinase (PTK) assay
The protein tyrosine kinase assay with cell lysates performed on PamChip arrays on the PamStation®12 is a flowthrough microarray assay to determine the activity of tyrosine kinases. For further information see supplemental material and methods.

Co-immunoprecipitation and subsequent mass spectrometry analysis
Co-immunoprecipitations were performed by using Dynabeads protein G (10003D, Invitrogen) according to the manufacturer's instruction in combination with IgG mouse (CS200621, Millipore) or anti-SFRP2 (sc-365524, Santa Cruz) antibodies. Further information about the coimmunoprecipitation conditions and subsequent mass spectrometry analysis can be found in supplemental material and methods.

Statistical Analysis
Data in bar graphs depict mean ± standard deviation. The laboratory experiments were repeated independently three times. Comparisons between the two groups were assessed using Student's t-tests (two-tailed) and χ 2 tests. p < 0.05 was considered to be statistically significant. Correlation rates were assessed by Pearson's correlation test.