Data-Driven Computational Modeling Identifies Determinants of Glioblastoma Response to SHP2 Inhibition

SH2 domain-containing phosphatase-2 (SHP2) is a proto-oncogene known for its ability to promote extracellular signal-regulated kinase (ERK) activity downstream of receptor tyrosine kinases (RTKs) (1, 2). While SHP2 is an attractive therapeutic target in cancer (3), it has been difficult to target with specificity until recently (4, 5). Allosteric SHP2 inhibitors such as SHP099 and TNO155 are now available and may help overcome adaptive resistance such as ERK reactivation after MEK inhibition in carcinomas (6). SHP2 targeting is also of longstanding interest for glioblastoma multiforme (GBM; refs. 7–9) due to largely ineffective standard of care [including temozolomide chemotherapy (10)] and involvement of SHP2 in RTK signaling commonly dysregulated in GBM (11). SHP2′s ubiquitous expression and broad functions may also make it an attractive target for overcoming issues of tumor heterogeneity and bypass signaling in GBM (12, 13). Although SHP2 inhibitors are in clinical trials, they remain untested in GBM.

While SHP2 antagonism may effectively target glioma stem cells (7) or PDGFR-driven GBM (14), the ultimate consequences of SHP2 inhibition may be challenging to foresee given that SHP2 can both promote and antagonize oncogenic signaling. In a previous study from our lab, SHP2 knockdown impeded GBM cell/tumor growth and hypoxia-inducible factor (HIF) expression but simultaneously promoted resistance to targeted therapy (15). These effects were attributed, in part, to the ability of SHP2 to drive ERK activity and concomitantly antagonize STAT3 phosphorylation. Other conflicting roles of SHP2 have been reported (e.g., refs. 16, 17). These complexities may be rooted in the ability of SHP2 to regulate a broader network of signaling pathways than has been characterized and in a way that depends on specific cell setting.

We hypothesized that SHP2 controls a network of GBM signaling processes that impact the effects of SHP2 inhibition in ways that depend on presence of other drugs, cell background, and tumor microenvironmental factors. Because we anticipated that the scope of SHP2-dependent signaling could be large, we adopted a data science approach wherein we broadly characterized signaling and phenotypic changes in GBM cells for a variety of perturbations and used the data to develop a quantitative multivariable regression model. Specifically, we constructed a partial least squares regression (PLSR) data-driven model based on dynamic measurements of 40 signaling nodes (protein phosphorylation or expression) for SHP2-depleted or -replete GBM cells treated with RTK inhibitors or temozolomide and cultured under normoxic or hypoxic conditions. Such data science approaches are increasingly used in cancer research because they can reduce the dimensionality of multivariate cell decision processes to a realistic number of actionable drug targets (18–20). The PLSR model we developed efficiently separated experimental conditions across three principal components, while exceeding accepted thresholds for model fit and predictive power. The model was used to develop three testable hypotheses, validation of which led to novel inferences about the ability of SHP2 inhibition to promote DNA damage repair or AKT bypass resistance to therapy and the ability of SHP2 to control HIF expression through multiple MAPKs. Validation in multiple GBM cell backgrounds and mouse tumor xenografts, and through analysis of publicly available patient data, suggests these phenomena may be relevant in the GBM clinical setting and lends support to biomedical data science models as useful tools for designing clinical trials.

U87MG cells expressing EGFRvIII (Dr. Frank Furnari, University of California San Diego, La Jolla, CA), U118MG (ATCC), T98G (ATCC), LN229 (ATCC), LN18 (ATCC), and LentiX 293T (Takara Bio) cells were maintained in DMEM with 10% FBS (VWR), 1 mmol/L l-glutamine, 100 U/mL penicillin, and 100 μg/mL streptomycin (Gibco). G88 (Dr. Jakub Godlewski, Harvard Medical School, Boston, MA), TS528 (Dr. Frank Furnari), T3691 (Dr. Celeste Simon, University of Pennsylvania, Philadelphia, PA), and T4302 (Dr. Celeste Simon) glioma stem cells (GSC) were maintained in suspension culture in Neurobasal Medium with B27 and N-2 supplements (Gibco), 0.25 mmol/L l-Glutamine, 100 U/mL penicillin, and 100 μg/mL streptomycin (Gibco) with 50 ng/mL recombinant human EGF (PeproTech) and human basic FGF (PeproTech). G88 cells expressing click beetle red luciferase (CBLuc) were described previously (21). Cell characteristics are listed in Supplementary Table S1. For hypoxic experiments, cells were cultured in 1% oxygen using a Forma Steri-cycle i160 tri-gas incubator. For chronic SHP2 inhibition, medium containing SHP099 was replenished every 2 to 3 days. U87MG and U118MG parental or transduced cell lines were tested for Mycoplasma using the MycoAlert system (Cambrex) and used within 15 passages from testing. G88 cells were authenticated by short tandem repeat profiling within six months of use.

Oligonucleotides for SHP2 shRNA (shRNA#1: 5′–GTATTGTACCAGAGTATTA–3′; shRNA#2: 5′–GGACGTTCATTGTGATTGA–3′) and control shRNA (5′–ATCACAGAATCGTCGTATGCA–3′) were cloned into the lentiviral pLKO.1-puro plasmid (Broad Institute, The RNAi Consortium), and vectors were used to generate U87MG, U118MG, T98G, and LN229 cells expressing shRNA. Identical oligonucleotides were cloned into TET-pLKO-puro (Dr. Dmitri Wiederschain, Novartis, Cambridge, MA; Addgene plasmid #21915) and used to generate U87MG cells expressing doxycycline-inducible shRNA. Identical oligonucleotides, or a control nontargeting sequence, were cloned into the lentiviral pSicoR-GFP plasmid (Dr. Tyler Jacks, Massachusetts Institute of Technology, Cambridge, MA) and used to generate G88 and TS528 GSCs expressing shRNA. Unless indicated otherwise, SHP2 shRNA#1 was used. The lentiviral pLKO.1-puro vector expressing HIF1α shRNA and pCDH-puro vector expressing HIF1α cDNA were gifts from Dr. Celeste Simon. The pBABE-hygro retroviral vector expressing PTEN was cloned previously (22) using cDNA provided by Dr. Frank Furnari.

Lentivirus was produced by calcium phosphate–mediated transfection of LentiX 293T cells with pLKO.1-puro, pSicoR-GFP, or pCDH-puro and packaging vectors pCMV-VSVg, pMDL-gp-RRE, and pRSV-Rev. Retrovirus was produced by calcium phosphate–mediated transfection of LentiX 293T with a pBABE-hygro plasmid and packaging vectors pCMV-VSVg and pUMVC.

MILLIPLEX Cell Signaling Multiplex Assay kits (Millipore Sigma) were used according to manufacturer's recommendations. Catalog numbers and analyte lists are provided in Supplementary Table S2.

Whole-cell lysates were prepared and protein concentrations determined as described for Luminex assays in Supplementary Materials and Methods. Approximately 20 μg of total protein was loaded per lane on 4% to 12% gradient polyacrylamide gels (Thermo Fisher Scientific) under denaturing and reducing conditions and transferred to 0.2 μm nitrocellulose membranes (Bio-Rad). After antibody probing, membranes were imaged on a LI-COR Odyssey CLx. Densitometry was performed using Image Studio (LI-COR). Antibody information is provided in Supplementary Materials and Methods.

RNA was extracted using the RNeasy Kit (Qiagen), and equal amounts were reverse transcribed (High Capacity cDNA Reverse Transcription Kit, Applied Biosystems). qRT-PCR was performed using SYBR Green PCR Master Mix (Applied Biosystems) on a StepOnePlus or QuantStudio3 Real-Time PCR System (Applied Biosystems). All qRT-PCR primers are listed in Supplementary Table S3.

Cells on glass coverslips were maintained in 6-well dishes, fixed in 4% paraformaldehyde in PBS for 20 minutes, permeabilized with 0.25% Triton X-100 in PBS for 5 minutes, and stained overnight at 4°C with primary antibodies. The next day, samples were washed with PBS and stained with secondary antibodies conjugated to Alexa Fluor fluorochromes (Thermo Fisher Scientific) and Hoechst (Thermo Fisher Scientific).

Flow cytometry measurements for cell death assays and cell-cycle analysis were performed on a BD Biosciences FACSCalibur at the UVA Flow Cytometry Core Facility.

For AP-1 reporter assays, cells were transfected with expression vector 3xAP1pGL3 and RSVpGL3 (Dr. Alexander Dent, Indiana University, Bloomington, IN; Addgene plasmids #40342, 40343).

Two-thousand G88 cells (parental or expressing control or SHP2 shRNA) were stereotactically injected into the right striatum of 4 to 8-week-old, female, BALB/c SCID mice (Charles River). After recovery, animals were randomly assigned to treatment groups (n ≥ 5 for all treatment groups; see figure legends for specific numbers for each experiment). Drugs preparations were: SHP099, 10 or 50 mg/kg in corn oil, oral administration; temozolomide, 15 mg/kg in sterile water, intraperitoneal injection. All experiments were approved by the UVA Institutional Animal Care and Use Committee and performed in accordance with NIH guidelines.

Charged slides with 4 to 5-μm–thick sections from formalin-fixed, paraffin-embedded tissues were generated by the UVA Research Histology Core. Antigen-retrieved tissues, prepared by the UVA Biorepository and Tissue Research Facility, were permeabilized with 0.10% Triton X-100 in PBS for 15 minutes, blocked for 1 hour at room temperature, and stained overnight at 4°C with antibodies at manufacturer-recommended dilutions. The next day, samples were washed and stained with Alexa Fluor–conjugated secondary antibodies and Hoechst (Thermo Fisher Scientific).

GBM patient data were downloaded from The Cancer Proteome Atlas portal (https://tcpaportal.org/tcpa/) and cBioPortal (https://www.cbioportal.org/) and analyzed as described in detail in Supplementary Materials and Methods.

Luminex data were normalized to baseline values (t = 0, DMSO, 21% oxygen) for control shRNA-expressing cells, individually for each replicate. Averages of normalized Luminex and phenotype measurements were mean-centered and variance-scaled (Supplementary Table S4). Partial least squares regression (PLSR) was performed using SIMCA-P (Umetrics). To improve the model, t = 0 time points, for which mean-centered and variance scaled values are identical for all analytes, and time points with qualitatively identical trends for a given analyte were removed from the X (signaling) matrix. A three-component model was chosen to maximize fit (R²X, R²Y) and predictive power (Q²) with the fewest principal components. The final three-component model met accepted criteria for model performance, with R²X and R²Y “goodness of fit” parameter values > 0.75 and a Q² “goodness of prediction” parameter value > 0.55, as in other cancer-specific models (23, 24). Supplementary Table S5 provides loadings for the first three principal components. For “filtered” PLSR loadings plots, only SHP2-dependent signals (P < 0.05 and log₂ fold changes > 0.5 or < −0.5 between control and SHP2 shRNA) were shown. Comparisons were made between control or SHP2 shRNA for (1) DMSO, 21% oxygen (2) G+P, 21% oxygen, (3) temozolomide, 21% oxygen, or (4) DMSO, 1% oxygen. See Supplementary Table S6 for P-value and fold change calculations for all comparisons. P values were determined using a two-tailed Student t test (Excel), followed by FDR adjustment by the Benjamini–Hochberg method (GraphPad).

Comparisons between two groups were made using a two-tailed Student t test (Excel), with significance determined as P < 0.05. Multiple comparisons were made using one-, two-, or three-way ANOVA (GraphPad), with P values calculated by Tukey post hoc testing and significance determined as P < 0.05. Unless stated otherwise, all data are representative of three independent replicates. Pearson correlation coefficients were calculated in Excel. P values for log-rank comparison tests of survival probability curves were calculated in the R package “survival.”

Additional details for the methods listed above, plus those for PTEN immunoprecipitation, subcellular fractionation, siRNA, and inhibitors/other reagents, are provided in Supplementary Information.

Constructing a data-driven model of SHP2 signaling control of GBM phenotypes requires paired measurements of cell phenotypes and signaling for a common set of conditions. We began with phenotype measurements. To characterize relationships among SHP2 expression and responses to drugs and hypoxia, an adherent GBM cell line expressing control or SHP2 shRNA was treated with gefitinib and PHA665752 (EGFR and MET inhibitors, respectively), temozolomide, or DMSO, and cultured in 21% or 1% oxygen. Measurements were made of cell-cycle distribution, cell death, and HIF1α expression (Fig. 1A–C; Supplementary Fig. S1A). U87MG cells expressing EGFR variant III (EGFRvIII), which is expressed in approximately 25% of GBM, were chosen for this dataset after screening three cell lines for two expected effects of SHP2 knockdown, reduced ERK phosphorylation and cell-cycle arrest (Supplementary Fig. S1B–S1D). EGFRvIII-expressing U87MG cells are hereafter referred to as U87MG for simplicity. Specificity of SHP2 knockdown effects was confirmed using a second nonoverlapping shRNA (Supplementary Fig. S1E–S1G).

Figure 1A–C demonstrate some expected trends based on prior work (15), including that SHP2 depletion promoted cell cycle arrest (Fig. 1A), increased resistance to EGFR and MET inhibitors (Fig. 1B), and impaired HIF1α expression (Fig. 1C). At the same time, additional trends were revealed that motivate a data-driven model. For example, the ability of SHP2 knockdown to promote resistance to temozolomide (Fig. 1B) may have important implications for combining SHP2 inhibitors with standard-of-care GBM chemotherapy.

To identify signaling processes responsible for the phenotypes in Fig. 1, we generated lysates from U87MG cells expressing control or SHP2 shRNA (plated in parallel) using the same conditions as in Fig. 1 and for t ≤ 48 hours after treatment (Fig. 2A). Lysates were analyzed using Luminex assay kits selected for coverage of pathways reported to be regulated by SHP2, including MAPK/SAPK, AKT/mTOR, STAT, RTK, apoptosis, and DNA damage signaling (25–30). Most Luminex antibodies targeted phosphorylated proteins. Signaling and phenotypic data were then preprocessed as described in Materials and Methods (Fig. 2B; Supplementary Table S4) and PLSR was performed. PLSR efficiently separated experimental conditions (SHP2 expression, drug treatments, hypoxia/normoxia), with >75% of the data variance explained by three principal components (PC) and sufficient predictive power (Q² > 0.55; ref. 31; Fig. 2C–E; Supplementary Fig. S2A; Supplementary Table S5). See Materials and Methods for additional details on model goodness of fit and prediction.

To focus on likely signaling consequences of targeting SHP2, we filtered the model to retain analytes that displayed a significant difference (P < 0.05) and log₂ fold change > 0.5 or < −0.5 between control and SHP2 shRNA conditions, with comparisons made for identical drug treatment, oxygen tension, and time (Fig. 2F; Supplementary Table S6). One obvious trend was that >80% of analytes projected positively in PC1, along with S and G₂ cell-cycle phases (Fig. 2G). Many top analytes in PC1 (e.g., EGFR, MEK/ERK, and JNK) have established relationships with SHP2 and regulate cell-cycle progression (25, 28). The direction of their projection suggested that their phosphorylation promotes cell-cycle progression. As expected, EGFR, MEK, and JNK inhibitors, as well as SHP2 inhibitor SHP099, antagonized cell-cycle progression (Fig. 2H). Similar effects of SHP099 were observed in GSCs (Supplementary Fig. S2B and S2C). In this broad sense, validated model predictions argue for the utility of SHP2 inhibitors in GBM, consistent with previous reports (7, 8, 14). The remainder of this study presents three specific and more nuanced inferences from the model relating to PCs 2 and 3.

In Fig. 2H, SHP2 inhibition produced cell-cycle effects similar to SHP2 knockdown (Fig. 1A). Based on that and projection of SHP2 knockdown conditions antiparallel to cell death in PC2 (Fig. 2C), we anticipated that SHP2 inhibition would generally antagonize cell death. However, acute SHP2 inhibition surprisingly promoted cell death (Fig. 3A; Supplementary Fig. S3A). We hypothesized that these differences could involve signaling adaptation with stable SHP2 knockdown. To pursue this, we generated U87MG cells expressing doxycycline-inducible control or SHP2 shRNA and, using them, observed time-dependent changes in cell death and ERK phosphorylation after doxycycline treatment (Fig. 3B; Supplementary Fig. S3B and S3C). Moreover, in differentiated GBM and GSC lines, chronic SHP2 inhibition (>4 days) promoted resistance to RTK inhibitors [Fig. 3C; Supplementary Fig. S3D; the marginally increased U87MG cell death in response EGFR and MET inhibition when combined with acute SHP099 (Fig. 3A) was lost with shorter treatment.] Furthermore, ERK inhibition in response to SHP2 inhibition was not durable, even with SHP099 replenishment (Fig. 3D; Supplementary Fig. S3E), again suggesting adaptation. Importantly, acute and chronic SHP2 inhibition and doxycycline-induced SHP2 knockdown all led to cell-cycle effects similar to constitutive SHP2 knockdown (Supplementary Fig. S3F–S3H).

While lack of durable ERK suppression could explain differences between acute and chronic SHP099 treatment, positive projection of ERK in PC2 (Fig. 2F) and evidence from our previous work (15) suggest that other SHP2-regulated pathways are relevant. To identify other pathways, lysates from cells treated for 1 (acute) or 10 (chronic) days with SHP099 were analyzed by Luminex (Fig. 3E). To further narrow candidates, we compared Fig. 3E with a version of the filtered PLSR model shown in Fig. 2F displaying only analytes that were significantly and substantially perturbed by SHP2 knockdown (P < 0.05 and log₂ fold change > 0.5 or < -0.5 between control and SHP2 shRNA conditions) when EGFR and MET inhibitors were present and focusing on analytes projecting antiparallel to cell death in PC2 (Fig. 3F). Although STAT5 meets the criteria, STAT5 and STAT3 phosphorylation increased in response to both acute and chronic SHP2 inhibition, suggesting they are unlikely responsible for the adaption. It is worth recalling that STAT3 was previously identified as promoting resistance to RTK inhibitors with stable SHP2 knockdown (15), and that relationship was observed here (Supplementary Fig. S3I and S3J).

The remaining analytes that met the criteria were predominantly from the AKT/mTOR pathway. As expected, based on projections of control and knockdown conditions in Fig. 2C, constitutive SHP2 knockdown promoted phosphorylation of AKT pathway analytes (Fig. 3G). Moreover, in response to doxycycline-induced SHP2 knockdown, AKT phosphorylation briefly decreased (2 days) and then increased above baseline (9 and 16 days; Fig. 3H; Supplementary Fig. S3K). Similar increases in AKT and GSK3β phosphorylation (without acute decrease) were observed in G88 GSCs treated with SHP099 (Fig. 3I). Interestingly, decreased AKT phosphorylation with acute SHP2 inhibition was only observed in a subset of differentiated GBM lines (Supplementary Fig. S3L), perhaps consistent with context-dependent SHP2 regulation of AKT (26). PI3K inhibition (GDC-0941) sensitized cells with SHP2 knockdown to EGFR and MET inhibition (Fig. 3J and K), confirming AKT's importance. Moreover, while PI3K inhibition did not augment response to gefitinib and PHA665752 in SHP099-naïve G88 cells, it promoted response to gefitinib and PHA665752 in cells chronically treated with SHP099 (Fig. 3L). Thus, PI3K and AKT inhibitors could be utilized to overcome resistance mechanisms induced by prolonged SHP2 inhibition.

We previously reported that HIF1α and HIF2α expression in GBM cells was impaired by SHP2 knockdown (15). This is important because hypoxia in the GBM microenvironment promotes tumor cell stemness and tumor progression (32, 33). While we previously attributed this effect to SHP2 control of ERK, the PLSR model implicated JNK, p38, and ERK MAPKs together, based on their projection in principal component 3 (Fig. 4A). We note that HIF2α expression mirrored HIF1α trends (Supplementary Fig. S4A), that ectopic HIF1α expression promoted cell-cycle arrest and resistance to RTK inhibition (Supplementary Fig. S4B–S4D), and that HIF1α antagonized resistance to RTK inhibitors in hypoxia (Supplementary Fig. S4E and S4F). Thus, HIF1α measurements captured more general HIF trends, and HIF1α may be directly responsible for SHP2-dependent phenotypes.

Returning to the PLSR model, MAPK/SAPK analytes projected more strongly in principal component 3 than others (Fig. 4A; Supplementary Fig. S4G), and ERK, JNK, and p38 phosphorylation decreased in response to SHP2 knockdown and hypoxia (Fig. 4B). Previous studies have suggested that hypoxia can positively or negatively regulate MAPKs, and that MAPKs can regulate HIF expression and transcriptional activity (34–36). To test the importance of individual MAPKs, we pretreated U87MG cells with MEK, JNK, p38, or SHP2 inhibitors, and evaluated HIF expression (Fig. 4C; Supplementary Fig. S4H). HIF1α and/or HIF2α expression in hypoxia was significantly impaired by all MAPK inhibitors. At the concentrations used, SHP099 reduced HIF1α and HIF2α expression more than any MAPK inhibitor, potentially because SHP099 more effectively inhibited ERK or because SHP2 inhibition targets multiple MAPKs. HIF1α expression at 1% O₂ was not influenced by PI3K or mTOR inhibition (Supplementary Fig. S4I).

In seeking to determine whether multiple SHP2-regulated MAPKs influence HIF expression, we noted that c-Jun, which is regulated by ERK, p38, and JNK (37, 38), projected strongly in PC3 (Fig. 4D) and that c-Jun has been implicated in HIF2α transcription in GBM (39). In differentiated GBM cells and GSCs, SHP2 knockdown or inhibition reduced HIF1/2α transcript abundance (Fig. 4E and F; Supplementary Fig. S4J and S4K). Moreover, the extent to which individual inhibitors reduced c-Jun phosphorylation correlated with HIF2α transcript abundance (Fig. 4G; Supplementary Fig. S4L). Interestingly, SB203580 has been reported to promote phosphorylation of ERK, JNK, and c-Jun (40), consistent with our observations (Supplementary Fig. S4H), and potentially explaining why SB203580 increased HIF2α transcripts. When cells were treated with proteasome inhibitor MG132, SHP2 knockdown–mediated differences in HIF1α expression were also apparent in normoxia (Supplementary Fig. S4M).

We further noted that nuclear c-Jun abundance (an indirect indicator of c-Jun transcriptional potential) correlated with nuclear HIF1α expression in hypoxia but was reduced by SHP2 inhibition (Fig. 4H). However, siRNA-mediated c-Jun knockdown did not decrease HIF1/2α transcripts (Fig. 4I). Given that c-Jun carries out its role in AP-1 complexes with other Jun or Fos family members (41), we screened cells with c-Jun knockdown for phosphorylation of another AP-1 member, fos-related antigen 1 (FRA1), to test for possible compensatory phosphorylation (Fig. 4J). Indeed, c-Jun knockdown, but not SHP2 inhibition, promoted FRA1 phosphorylation. These data suggest that SHP2-dependent differences in HIF1/2α transcripts may be generally due to AP-1 complexes. Importantly, SHP2-regulated MAPKs regulate multiple AP-1 complex members (42). Furthermore, promoters of HIF1A and EPAS1 (which encodes HIF2α) contain AP-1 binding sites, a subset of which preferentially bind Jun/Fos heterodimers (Supplementary Fig. S4N). Using an AP-1 DNA-binding reporter, we observed that AP-1 transcriptional activity decreased with SHP2 knockdown, consistent with trends in HIF1/2α transcripts (Fig. 4K). Similarly, AP-1 transcriptional activity in response to SHP2 or MAPK inhibitors mirrored HIF2α transcript levels (Fig. 4L; Supplementary Fig. S4L). A qPCR array revealed broader SHP2 regulation of hypoxia-regulated genes (Fig. 4M). SHP2 inhibition also influenced transcriptional response to hypoxia, as demonstrated by VEGF expression (Supplementary Fig. S4O).

In Fig. 1, we learned that SHP2 antagonism promotes resistance to temozolomide. To understand this, we generated a version of the filtered PLSR model shown in Fig. 2F displaying only analytes that were significantly and substantially perturbed by SHP2 knockdown (P < 0.05 and log₂ fold change > 0.5 or < −0.5 between control and SHP2 shRNA conditions) when temozolomide was present (Fig. 5A). Recalling that several AKT/mTOR analytes projected antiparallel to cell death when screening based on response to RTK inhibitors, it was striking that a single AKT/mTOR pathway member, PTEN, strongly projected antiparallel to cell death for temozolomide (Supplementary Fig. S5A). Inspection of the Luminex data used to generate the PLSR model revealed that temozolomide promoted PTEN S380 phosphorylation, which was augmented by SHP2 knockdown (Fig. 5B). While PTEN is best known as a lipid phosphatase, it can also participate in DNA damage repair (43). Phosphorylation of PTEN S380, and other C-terminal serine/threonine residues, can promote PTEN nuclear translocation and participation in DNA damage repair by facilitating RAD51 chromatin loading (44–47). Interestingly, PTEN S380 phosphorylation correlated best with AKT S473 and GSK3β S9 phosphorylation for control treatment (Pearson correlation coefficients of 0.615 and 0.605, respectively). In the presence of temozolomide, correlation coefficients of the DNA damage indicators CHK1 S345 phosphorylation and MDM2 expression (0.319 and 0.289, respectively) exceeded those for AKT S473 and GSK3β S9 (0.017 and 0.059, respectively; Fig. 5C).

To test the importance of PTEN, we transfected cells with control or PTEN-targeting siRNA and measured temozolomide-induced cell death (Fig. 5D; Supplementary Fig. S5B). In U87MG (PTEN exon-deletion mutant retaining phosphatase activity; ref. 48) and LN18 cells (wild-type PTEN), PTEN knockdown promoted AKT phosphorylation but also promoted sensitivity to temozolomide. In both cell lines, ectopic wild-type PTEN promoted resistance to temozolomide (Fig. 5E). Moreover, SHP2 inhibition promoted resistance to temozolomide in differentiated GBM lines and GSCs with intact PTEN expression (Fig. 5F; Supplementary Fig. S5C), but did not alter response to temozolomide in T3691 GSCs, which express little PTEN (Supplementary Fig. S5D). In LN18 cells, SHP2 inhibition promoted PTEN S380 phosphorylation and reduced PARP cleavage and H2A.X phosphorylation (Fig. 5G; Supplementary Fig. S5E). SHP2 knockdown also promoted PTEN nuclear localization in response to temozolomide (Fig. 5H and I; Supplementary Fig. S5F).

Regulation of PTEN serine phosphorylation by SHP2 (a tyrosine phosphatase) is likely indirect, but we hypothesized that relevant PTEN tyrosines could be phosphorylated in response to DNA damage. PTEN phosphorylation at Y240 by fibroblast growth factor receptor (FGFR) has been implicated in GBM radioresistance (49), and Y240 and Y315 reside in the PTEN C2 domain, which regulates membrane binding (50). We suspected that their phosphorylation could promote PTEN membrane dissociation, as observed with C-terminal serine phosphorylation. In fact, in a prior study of radiation-induced DNA damage, tyrosine-phosphorylated PTEN was detected exclusively in the nucleus, and Y240F PTEN mutants displayed impaired nuclear localization (49). Using ScanSite (51), we identified EGFR and PDGFR as possible drivers of PTEN Y240 and Y315 phosphorylation (Supplementary Fig. S5G). We previously identified EGFRvIII as a possible SHP2 substrate (15), and others have shown that SHP2 can dephosphorylate RTKs (52). In addition, our data showed that SHP2 knockdown promoted EGFR, FGFR1, and PDGFRβ phosphorylation (Supplementary Fig. S5H). Because we observed that temozolomide promoted EGFR phosphorylation in G88 cells (Supplementary Fig. S5I), we hypothesized that EGFR could be an SHP2-regulated RTK that promotes PTEN tyrosine phosphorylation. Consistent with this, SHP2 knockdown promoted EGFR and PTEN tyrosine phosphorylation and survival in cells treated with temozolomide (Fig. 5J; Supplementary Fig. S5J–S5L). In SHP2 knockdown cells, EGFR inhibition prevented temozolomide-induced PTEN nuclear localization (Fig. 5K; Supplementary Fig. S5M) and restored cell death response to temozolomide to levels observed in SHP2-replete cells (Fig. 5L), consistent with the proposed mechanism (Fig. 5M).

To assess response to SHP099 in vivo, G88 cells expressing luciferase were orthotopically implanted in BALB/c SCID mice. Longitudinal bioluminescence imaging of mice treated with SHP099 (daily, 10 or 50 mg/kg, 5 days after implantation) revealed a dose-dependent decrease in tumor size and disease burden (determined by overall weight loss) in response to SHP099 (Fig. 6A; Supplementary Fig. S6A and S6B). Expected differences in SHP2 Y542 phosphorylation (53), a proxy for SHP2 activity (1), and percentage of cells expressing Ki67 (Fig. 6B; Supplementary Fig. S6C) were observed.

To interrogate relationships between SHP2 and HIF, we identified tissue void regions resembling necrotic zones found in human GBM (54) that contained most HIF2α-positive nuclei (Fig. 6C and D). Nuclear HIF2α expression was decreased in tumors treated with SHP099 (Fig. 6D). Consistent with in vitro observations, SHP099 reduced nuclear c-Jun abundance (Fig. 6E) and VEGF transcripts (Fig. 6F) and elevated EGFR phosphorylation (Fig. 6G; Supplementary Fig. S6D).

To investigate SHP2 effects on temozolomide response, G88 cells expressing control or SHP2 shRNA were used in an orthotopic model (daily temozolomide, starting 10 days after implantation; see Supplementary Fig. S6E). After 3 days of temozolomide, MRI revealed that tumors expressing SHP2 shRNA were smaller but less responsive than controls (Fig. 6H; Supplementary Fig. S6F). Tumor sizes were consistent with overall disease burden (Supplementary Fig. S6G). In fact, mice with SHP2 shRNA tumors treated with temozolomide lost as much weight as untreated controls. Ex vivo analyses supported MRI-based comparisons of tumor size (Fig. 6I; Supplementary Fig. S6H and S6I), and revealed decreased ERK phosphorylation with SHP2 knockdown (Supplementary Fig. S6J). Consistent with the proposed mechanism of temozolomide resistance, SHP2 knockdown tumors displayed elevated EGFR phosphorylation, elevated nuclear PTEN, and decreased H2A.X phosphorylation in response to temozolomide (Fig. 6I and 6J). Tumor sizes were consistent with Ki67 staining (Supplementary Fig. S6K).

To further explore SHP2 regulation of DNA damage response, we analyzed The Cancer Genome Atlas (TCGA) data by pairing survival outcomes with phosphorylated SHP2 (Y542) measurements from reverse phase protein arrays (55). While SHP2 phosphorylation was not prognostic for all GBM patients (P = 0.22 by log-rank test comparing survival probabilities; Supplementary Fig. S6L), a potential survival difference did emerge for patients treated with temozolomide (P = 0.057; Fig. 6K). Temozolomide-treated patients whose tumors exhibited high (above median) SHP2 phosphorylation had a median survival of 21.9 months, compared with 13.0 months for those with low (below median) SHP2 phosphorylation. Furthermore, tumors with elevated SHP2 Y542 phosphorylation displayed elevated c-Jun phosphorylation and expression of CD31 (tumor vascularization marker) and VEGF (Supplementary Fig. S6M), consistent with other model inferences described in Results.

The data-driven model developed here predicted multiple SHP2-dependent signaling events that influence SHP2 inhibitor efficacy in GBM (Fig. 7). Predictions developed from a single cell line were validated in multiple GBM cell lines, with differences among cell lines reinforcing model-predicted relationships (e.g., effects of PTEN expression). While SHP2 inhibitor responses can be studied purely experimentally (e.g., ref, 53), data-driven models, which have been used to study effects of kinase inhibitors (20, 56), may be particularly useful for understanding multivariate effects of phosphatase inhibitors, given the substrate promiscuity of phosphatases (57). Such models can also directly inform clinical trial design, given that our findings can be readily incorporated into single-agent and combination trials for patients with GBM. Whether inferences from our study are applicable outside the GBM context remains to be seen.

One limitation revealed by pursuing model inference (I) was the model's inability to predict qualitatively different effects of short- versus long-term SHP2 inhibition on AKT signaling. Such limitations can potentially be overcome by tracking both signaling and phenotypes over time. The substantial increase in data dimensionality of such problems can be efficiently addressed by tensor PLSR (58). It may also eventually be possible to integrate data-driven models with mechanistic models of SHP2 regulation (59) to predict the impact of targeting proteins involved in SHP2 regulation on cancer phenotypes.

The finding that SHP2 inhibition impaired HIF expression in GSC-derived orthotopic tumors has important implications given the roles of HIFs in GSC self-renewal and tumorigenic potential (60) and ongoing efforts to therapeutically target HIFs (e.g., trial of HIF2α inhibitor PT2977, NCT02974738). If SHP2 broadly controls HIF expression through effects of multiple MAPKs on AP-1 complex members, as our data suggest, SHP2 inhibitors may be better candidates for targeting hypoxia response than individual kinase inhibitors (NCT01497444; ref. 61). While our studies focused on SHP2 control of AP-1, nontranscriptional mechanisms may also be relevant. Indeed, MAPKs can phosphorylate HIFs in ways that impact protein stability and transcriptional activity (62, 63). Ultimately, for inhibitors of SHP2 and HIFs, evaluation in immunocompetent GBM models will be necessary as both proteins have immune cell roles (64, 65). In fact, an ongoing phase Ib clinical trial of SHP2 inhibitor TNO155 in solid tumor malignancies includes combination treatment with PD1 mAb spartalizumab (NCT04000529).

The unanticipated effects of SHP2 inhibition on GBM response to temozolomide suggest a potential need for tumor molecular profiling in deploying SHP2 inhibitors. Indeed, our findings could translate directly to a clinical trial combining an SHP2 inhibitor with temozolomide in patients with GBM, with PTEN status indicated by genetic profiling. Results of data-driven studies such as these will be increasingly applicable clinically, given growing availability of genetic profiling of GBMs and other cancers through Foundation Medicine and other providers. The effects we observed appeared independent of other indicators of temozolomide response, such as MGMT promoter methylation, based on characteristics of the cells studied (Supplementary Fig. S5C; Supplementary Table S1). Moreover, SHP2 inhibitors may broadly antagonize DNA damaging therapies, given that PTEN-dependent DNA damage repair in GBM was originally identified with ionizing radiation (49). Interestingly, SHP099 plus temozolomide has been reported to be an effective combination in mouse orthotopic GBM xenografts (14). One potential explanation is the unknown PTEN-status of the tumor model. Ultimately, further preclinical evaluation of combination therapies involving SHP2 inhibitors, potentially using staggered or periodic dosing to minimize undesirable consequences of simultaneous or continuous treatments, will be needed to maximize therapeutic benefit of SHP2 inhibitors for patients with GBM.

E.K. Day reports grants from National Science Foundation during the conduct of the study and since the first submission of this manuscript. E.K. Day began a postdoctoral position at Novartis (Cambridge, MA). Allosteric SHP2 inhibitors, such as SHP099 used in this study, are currently being developed by Novartis. M.J. Lazzara reports grants from National Science Foundation, American Cancer Society, and grants from National Cancer Institute during the conduct of the study. No disclosures were reported by the other authors.

E.K. Day: Conceptualization, funding acquisition, investigation, writing–original draft, writing–review and editing. Q. Zhong: Investigation. B. Purow: Resources, writing–review and editing. M.J. Lazzara: Conceptualization, funding acquisition, writing–original draft, writing–review and editing.

We thank UVA core facility staff for technical assistance, Dr. Frank Furnari for the pY240-PTEN antibody, Sara Adair for technical support, and Dr. Marieke Jones for assistance with statistics. This work was supported by NSF CBET 1511853 (to M.J. Lazzara), American Cancer Society RSG-15-010-01-CDD (to M.J. Lazzara), NSF GRFP DGE-1321851 (to E.K. Day), and UVA Cancer Center Support Grant from the NCI (P30CA044579).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.