A Genome-scale CRISPR Screen Identifies the ERBB and mTOR Signaling Networks as Key Determinants of Response to PI3K Inhibition in Pancreatic Cancer

The 10-year survival rate for pancreatic ductal adenocarcinoma (PDAC) has remained at just 3% for the past 40 years (1). Activation of the oncogene, KRAS, is one of the earliest genetic alterations detected in the development of PDAC and KRAS mutations are found in over 90% of cases (2, 3). Transgenic mouse models expressing oncogenic Kras^G12D have demonstrated that mutant KRAS is an important driver in pancreatic cancer, as switching off Kras signaling results in tumor regression (4). Recently discovered inhibitors of KRAS^12C have further validated the dependency of pancreatic, colon, and lung tumor models on oncogenic KRAS and demonstrated promising early clinical activity (5, 6).

There is a strong rationale for targeting PI3K in PDAC. Activation of the PI3K pathway is commonly observed in PDAC patient samples (7–9), regulates cell metabolism, growth, and survival, and is commonly implicated as a driver of human cancer (10). Importantly, phosphorylation of PI3K signaling markers, including AKT (Ser473), mTOR (Ser2448), and GSK3β (Ser9; ref. 11) or low expression of PTEN, a suppressor of PI3K signaling (12), is predictive of poor survival in pancreatic cancer. Moreover, the interaction between Ras and PI3Kα is essential for Kras^G12D-induced tumorigenesis in mice (13). Notably, Kras^G12D-driven murine PDAC tumors are dependent on PI3Kα (14, 15), but not PI3Kβ (15), or Craf (14) for tumorigenesis. Consequently, PI3K signaling is an attractive therapeutic target for PDAC. However, clinical trials of allosteric mTOR inhibitors, including temsirolimus (7), or everolimus (16), have shown limited activity in patients with gemcitabine-refractory, metastatic pancreatic cancer, likely due to loss of negative feedback on IRS1 and reactivation of PI3K (16). Multiple oncogenic pathways are engaged downstream of KRAS (17, 18), so it is perhaps unsurprising that targeting a single downstream effector may not be enough to affect cell viability. We hypothesize that PI3K inhibition selects for compensatory mechanisms sufficient to maintain tumor cell survival.

This study aimed to elucidate the mechanisms of intrinsic resistance to PI3K inhibition in pancreatic cancer and identify rational drug combinations to overcome them. Functional genomic screens have successfully identified loss-of-function events that drive drug resistance, finding NF1 loss to be a key driver of resistance to RAF inhibition in melanoma (19). We therefore employed a genome-scale synthetic lethal CRISPR screen to find loss-of-function events that could modulate sensitivity to PI3K inhibition. We discovered that the ERBB and mTOR signaling networks regulate response to PI3K inhibition in PDAC. Furthermore, we used a combination drug screen to prioritize clinically relevant targeted agents that synergize with PI3K inhibition to improve therapeutic response.

Pancreatic cancer cell lines were a kind gift from Dr. Anguraj Sadanandam (The Institute of Cancer Research, London, United Kingdom), with the exception of PANC1, PATU8902, MIAPACA2, YAPC, and HEK293T cells, which were obtained from the ATCC. T47D cells were from the Deutsche Sammlung von Mikroorganismen und Zellkulturen (DSMZ). All cells were cultured in DMEM (Sigma) supplemented with 10 % FBS (FBS Good, Pan Biotech), with the exception of MIAPACA2, which was supplemented with 20% FBS. Human pancreatic stellate cells (PSC) were obtained from ScienCell laboratories. Recombinant growth factors were obtained from Bio-Techne. Cell lines were tested for Mycoplasma using the MycoAlert Mycoplasma Detection Kit (Lonza). Cell line authentication was not performed.

All small-molecule inhibitors were purchased from Selleck Chemicals: BYL719 (S2814), GDC0941 (S1065), pelitinib (S1392), everolimus (S1120), AZD8055 (S1555), AZD2014 (S2783), and BEZ235 (S1009). Stock solutions were prepared in dimethyl sulfoxide (DMSO) and stored at −20°C.

For GI₅₀ determination, cells were seeded in 96-well plates. The next day, cells were treated with increasing concentrations of inhibitor or with DMSO alone. After a 72-hour incubation period, cell proliferation was quantified using CellTiter-Blue reagent (Promega) and normalized to DMSO-treated wells. GI₅₀ values were calculated using nonlinear regression analysis in GraphPad Prism software. For population doubling experiments, cells were seeded at an initial density of 1 × 10⁷ cells/flask in 225 cm² flasks. Cells were allowed to proliferate to 80%–90% confluence before they were counted and then reseeded at the same initial density. Population doublings (PD) were calculated according to the equation below.

For determination of maximum excess above bliss, cells were treated with a matrix of increasing concentrations of two inhibitors or DMSO. After a 72-hour incubation period, cell proliferation was quantified using CellTiter-Blue reagent and normalized to the DMSO-treated well. The Bliss independence model (20) was used to calculate synergy.

For colony assays, cells were seeded in 12-well plates. The next day, triplicate wells were treated with DMSO, the inhibitors alone, or the combinations. After 14 days, cells were washed with PBS and fixed in 4 % formaldehyde/PBS for 30 minutes. Cells were stained with 0.5% crystal violet in 70% ethanol and imaged using a FluoroChem E imaging system (Protein Simple). Colonies were quantified by solubilizing the crystal violet solution in 10% acetic acid and reading the absorbance at 595 nm using an EMax Plus Microplate Reader (Molecular Devices).

Human PSCs were cultured in DMEM/Nutrient Mixture F-12 Ham (Sigma Aldrich), supplemented with 1% GlutaMAX (Thermo Fisher Scientific), 1% Amphotericin B (Thermo Fisher Scientific), 1% Penicillin-Streptomycin (Sigma Aldrich), and 10% FBS (FBS Good, Pan Biotech). Cells were seeded in coculture with established PDAC cell lines at a starting density of 1 × 10³ cells/well in 96-well Ultra-low Attachment Round Bottom Multi-well Plates (Nexcelom). Cells were seeded to form spheroids at a ratio of 1:1 PSCs to PDAC cell lines. The next day, cells were treated with DMSO, fixed concentrations of drugs or the desired combinations. Spheroid diameter was measured over a time period of 10 days, with measurements taken every 3–4 days. The first measurement was taken the day after cells were plated, before the addition of DMSO and drug treatments. Spheroid diameter was imaged and quantified using the Celigo Imaging Cytometer (Nexcelom) and is the average of at least three replicate spheroids. For viability staining, spheroids were incubated with 1 μmol/L calcein AM and 40 μg/mL propidium iodide for 30 minutes prior to imaging.

After the desired treatment, cells were washed with cold PBS and lysed in NP40 buffer [0.5% NP40, 150 mmol/L NaCl, 50 mmol/L Tris pH 7.5, Pierce Protease and Phosphatase Inhibitor Mini Tablets (Life Technologies)]. Where detection of KRAS was necessary, cells were lysed in SDS buffer (1 % SDS, 10 mmol/L EDTA, 50 mmol/L Tris, pH 8). Bicinchoninic acid (Sigma) was used to determine protein concentration. Equal amounts of protein were separated by gel electrophoresis, using NuPAGE polyacrylamide gels (Life Technologies). Proteins were transferred to a nitrocellulose membrane using the iBlot 2 system (Life Technologies) and then blocked with LI-COR blocking buffer (LI-COR Biosciences). Membranes were incubated with the primary antibodies overnight at 4°C, followed by IRdye-conjugated secondary antibodies (LI-COR Biosciences) and detected using an Odyssey Fc imaging system (LI-COR Biosciences). Quantification of Western blots was performed using Image Studio Lite (LI-COR Biosciences). Details of the antibodies used can be found in Supplementary Table S1.

HEK293T cells were seeded at a density of 2.4 × 10⁶ cells/plate in 10-cm plates. The next day cells were transfected with shRNA/sgRNA plasmid (3 μg) and the packaging plasmids psPAX2 (2.1 μg) and pmD2.G (0.9 μg) using 30 μL lipofectamine per transfection. Cells were incubated for 72 hours at 37°C, after which the supernatant was collected and stored in 0.5 mL aliquots at −80°C for future experiments. Each batch of lentivirus was titrated on cells to determine concentration needed for 100% infection efficiency.

MISSION shRNA plasmids (pLKO.1) were obtained from Sigma-Aldrich. The pLKO.1-LacZ and -Luciferase targeting shRNA plasmids were from the Genetic Perturbation Platform (The Broad Institute). TRC numbers and target sequences for shRNA plasmids are shown in Supplementary Table S2. Cells were transduced with lentivirus as described previously (19). Cell proliferation was quantified using CellTiter-Blue reagent (Promega) and normalized to cells transduced with control lentivirus. Gene dependency scores were calculated on the basis of the dependency index described by Singh and colleagues (21).

Cells were plated in 384-well plates and the Echo 550 Liquid Handler (Labcyte) was used to dispense 20 nL of each compound from a library of 485 FDA-approved drugs and tool compounds (selected by the Cancer Research UK Cancer Therapeutics Unit and purchased from Selleckchem) onto the plates to give the final concentration of 800 nmol/L on the cells. Plates were then treated with either 100 nL of DMSO, BYL719, or GDC0941, to give a final concentration of 10 μmol/L BYL719 or 1 μmol/L GDC0941. After a 96-hour incubation period, cell proliferation was quantified using CellTiter-Blue reagent (Promega). Synergy was calculated using the Bliss independence model, as described previously. The Bliss score for each combination is the mean of three replicates.

LentiCRISPRv2 (was a gift from Feng Zhang, Addgene plasmid #52961; ref. 22) were digested with Esp3I at 37°C overnight (New England Biolabs, NEB). Oligos were designed to include each sgRNA target sequence (Supplementary Table S3) according to the “Zhang Lab General Cloning Protocol,” available at https://www.addgene.org/crispr/zhang/. Each pair of oligos was phosphorylated and annealed with T4 PNK enzyme (NEB). Each oligo duplex was then ligated into the appropriate vector using the quick ligase enzyme (NEB) at 16°C overnight. Lentiviral plasmids were transformed into Stbl3 bacteria (Invitrogen), according to the manufacturer's instructions, and then plated on ampicillin (50 mg/mL)-selective agar plates. Single colonies were then amplified, extracted, and used to produce lentivirus as described above. Before generation of lentivirus, each amplified plasmid was sequenced to ensure successful sgRNA sequence integration. To generate clonal cell populations expressing each plasmid, cells were first transduced with the virus. Cells were transduced with the lentiCRISPRv2 lentivirus, as this also contains the vector for Cas9 expression. Cells successfully transduced with lentiCRISPRv2 were selected for using 10 μg/mL blasticidin, respectively. After 7 days of selection, cells were seeded in 96-well plates at a density of 0.5 cells/well to select for clonal populations. These were expanded under continued antibiotic selection until sufficient cell numbers were generated. Stocks were frozen down in FBS with 10% DMSO and stored in liquid nitrogen.

Cell lines were engineered to express Cas9 by centrifugation of 4 × 10⁶ cells with (pXPR101) Cas9 lentivirus (1:1), in the presence of 8 μg/mL polybrene for 1 hour at 37°C. Cells were incubated with fresh media overnight, before cells were trypsinized and pooled for selection. Cells were incubated with 10 μg/mL blasticidin for 7 days to select for successfully infected cells. In parallel, cells were plated in 6-well plates for determination of infection efficiency. To assess Cas9 activity, parental and Cas9-expressing cells were infected with a lentivirus encoding both EGFP and a sgRNA targeting EGFP (pXPR_011-sgEGFP). Successfully transduced cells were selected for using 2 μg/mL puromycin, until all cells of a “no infection control” were dead. Cells were assayed by flow cytometry to assess EGFP expression. The activity of Cas9 was taken as the proportion of EGFP-negative cells in the Cas9-transduced population.

Cells were seeded in 12-well plates at a density of 3 × 10⁶ cells/well in 2 mL media. Cells were infected with the Avana4 lentiviral library (Broad Institute, 74,687 sgRNAs targeting 18,407 genes; refs. 23, 24) in four infection replicates. Cells were infected with a predicted representation of 500 cells/sgRNA after selection and with the volume of virus/well that gave approximately 40% infection efficiency. Cells were centrifuged at 2,000 rpm for 2 hours at 30°C in the presence of lentivirus and 8 μg/mL polybrene, followed by incubation in fresh media overnight. Cells were pooled and seeded into T225 flasks at a density of 1 × 10⁷ cells/flask for selection with 2 μg/mL puromycin for 7 days and passaged as necessary. In parallel, cells were seeded in 6-well plates to determine infection efficiency. After 7 days of selection, MIAPACA2 cells were split into three arms and treated with either 0.02% DMSO, 10 μmol/L BYL719, or 1 μmol/L GDC0941. Throughout the screen, cells were passaged as necessary, maintaining a total representation of 500 cells/sgRNA in each replicate. After eight population doublings, cells from each arm were collected and cell pellets stored at −80°C. Genomic DNA was extracted using the QIAamp DNA Blood Maxi Kit (Qiagen). PCR amplification and next-generation sequencing (NGS) was conducted as described previously (23). Briefly, sequencing adaptors and sample barcodes were added to sgRNA sequences from gDNA and pDNA samples by PCR. Samples were purified with Agencourt AMPure XP SPRI beads (Beckman Coulter A63880) and then sequenced on a HiSeq2000 (Illumina). Reads were counted by searching for the CACCG sequence of each sgRNA insert and then mapping the 20 nucleotide sgRNA sequence to a reference file of all sgRNAs in the library and assigned to the treatment condition based on the appended barcode.

The custom minipool lentiviral library [3,067 sgRNAs targeting 296 top hit genes (one gene was accidentally omitted), 496 nontargeting control sgRNAs, and 201 sgRNAs targeting essential genes] was prepared as described previously (23, 24). Plasmid DNA (pDNA) was sequenced by NGS to quantify the abundance of each sgRNA in the pool. The pDNA pool was then transfected into HEK 293T cells to produce lentivirus according to the “Large Scale Lentiviral Production” protocol available at https://portals.broadinstitute.org/gpp/public/resources/protocols. Each cell line was infected with the custom minipool lentiviral library in four infection replicates. Cells were infected with a predicted representation of 2,000 cells/sgRNA, after selection, and with the volume of virus/well that gave approximately 40% infection efficiency, as determined previously. The transduction and selection protocol used was the same as in the genome-wide screen and is described above. After 7 days of selection, cells were split into three arms and treated with either 0.02% DMSO, BYL719, or GDC0941. Throughout the screen, cells were passaged to maintain a representation of 2,000 cells/sgRNA. After eight population doublings, cells from each arm were collected and genomic DNA was extracted and sequenced, as in the whole-genome screen.

The abundance of each sgRNA in each replicate was quantified by calculating the Log₂(sequencing reads/million) (RPM), according to the formula below.

The log₂ fold change (LFC) from the early time point was calculated by normalizing RPM for each sgRNA in each replicate to that in an early time point control taken 3 days after selection with puromycin. The LFC between the DMSO and drug-treated arms was calculated as the difference in average LFC across three replicates. This was used to rank individual sgRNAs according to their selective depletion or enrichment in the drug-treated arms. Top scoring genes were ranked according to the number of independent high scoring sgRNAs targeting the same gene, according to the STARS gene-ranking algorithm (23). To assess depletion of essential genes from the population, as a positive control for successful gene editing, the RPM for each sgRNA was normalized to the plasmid DNA to calculate the LFC from baseline. The list of 885 core essential genes was kindly provided by Dr. Marco Licciardello (The Institute of Cancer Research, London, United Kingdom) and is compiled from the genes that were consistently and significantly depleted in all cell lines tested from three previous publications (25–27). To assess the statistical significance of the overlap between genes that modulated the response to BYL719 or GDC0941, the representation factor was calculated as below.

The associated probability was calculated by exact hypergeometric probability as detailed

(http://nemates.org/MA/progs/representation.stats.html).

A list of all hit genes was compiled from those that were that were enriched or depleted from the BYL719- or GDC0941-treated arms of the genome-wide CRISPR screen (FDR < 0.3). This list was used to interrogate the Reactome (28) and KEGG (29, 30) gene sets within the Molecular signatures database (31, 32) available at http://software.broadinstitute.org/gsea/msigdb.

To discover loss-of-function events that modulate sensitivity to PI3K inhibition in PDAC, we conducted a genome-wide CRISPR screen anchored to the PI3Kα-selective inhibitor BYL719 (alpelisib; ref. 33) or the pan-class I PI3K inhibitor GDC0941 (pictilisib; ref. 34). In vitro, pancreatic cell lines were resistant to BYL719 and GDC0941, compared with the PI3Kα-dependent breast cancer cell line, T47D (Fig. 1A; Supplementary Fig. S1A; ref. 35). Resistance was observed despite inhibition of PI3K–AKT signaling, highlighting a disconnect between pathway inhibition and inhibition of cell proliferation in the pancreatic cells (Fig. 1B; Supplementary Fig. S1E). We selected MIAPACA2 cells for the genome-wide screen, as they were resistant to both BYL719 and GDC0941 and dependent on KRAS for proliferation (Supplementary Fig. S1A–S1C). We engineered this cell line to stably express Cas9 (MIAPACA2_Cas9) and confirmed that Cas9 expression did not alter the response to PI3K inhibition by BYL719 or the pan-PI3K inhibitor GDC0941 (Supplementary Fig. S1D and S1E).

MIAPACA2_Cas9 cells were transduced with Avana4 lentiviral library (23, 24). Cells were split into DMSO-, BYL719-, or GDC0941-treated arms and passaged for eight population doublings (Fig. 1C). Cell proliferation rate was slowed by incubation with 10 μmol/L BYL719 or 1 μmol/L GDC0941 (Fig. 1D), concentrations at which PI3K signaling was near-completely inhibited. Our choice of inhibitor concentration was driven by a balance of attaining near-complete pathway inhibition at a concentration that permits sufficient cell proliferation for the screen to be performed within approximately 3–4 weeks. Although the concentration of BYL719 used was quite high, we hypothesized that by using two chemically distinct inhibitors, off-target effects could be accounted for by focusing on hits observed with both inhibitors. Cells were harvested, genomic DNA (gDNA) extracted, and sgRNAs amplified and barcoded by PCR. Next-generation sequencing (NGS) was employed to quantify the abundance of each sgRNA in each experimental arm. This analysis demonstrated that sgRNAs targeting essential genes or KRAS, a known dependency of this cell line, were depleted from the population, whereas nontargeting sgRNAs were not (Supplementary Fig. S2A). The STARS algorithm (23, 24) was used to rank genes with multiple scoring sgRNAs that were selectively depleted (enhancers of the antiproliferative effect) or enriched (suppressors of the antiproliferative effect) in each arm of the screen compared with DMSO (Fig. 1E and F; Supplementary Fig. S2B and S2C; Supplementary Tables S4–S7).

Overall, there was a large degree of overlap among the hit genes [false discovery rate (FDR) of <0.3] for which sgRNAs were depleted or enriched with BYL719 or GDC0941 treatment (Fig. 1G; Supplementary Table S8). Out of 82 and 34 genes for which sgRNAs enhanced the antiproliferative effect in the BYL719- and GDC0941-treated arms, respectively, 11 of the hits were common to both treatment arms (representation factor 75, P < 1.192 × 10⁻¹⁸). There was a greater degree of overlap between the genes for which sgRNAs were enriched with drug treatment, with 63 genes common to both drugs, out of a total of 135 and 120 genes that were enriched with BYL719 and GDC0941, respectively (representation factor of 74, P < 6.147 × 10⁻¹⁰⁹). Strikingly, sgRNAs targeting multiple negative regulators of PI3K/mTOR signaling were enriched in the drug-treated populations, implying that loss of these genes promotes resistance to PI3K inhibition. Indeed, the top-ranking sgRNAs that were enriched in both drug-treated arms targeted TSC1 and TSC2, which inhibit the activity of RHEB and downstream mTORC1 signaling (36). Other sgRNAs enriched under drug treatment targeted PTEN, DDIT4 (36), AKT1S1 (37), and RALGAPB (38). Guide RNA–targeting genes encoding proteins of the mTORC1 network, mTOR kinase, RAPTOR, and RRAGC were significantly depleted from the BYL719-treated arm, suggesting that loss of mTORC1 sensitizes cells to BYL719 treatment (39, 40). This implicates the mTORC1 complex as a key mediator of resistance to PI3K inhibition. Given that loss of PTEN, TSC1, or TSC2 confers resistance to PI3K inhibition and loss of mTOR sensitizes to PI3K inhibition, this provides important validation of the screening conditions, as these events are known to modulate sensitivity to PI3K inhibitors (41, 42).

We used the Molecular Signatures Database (MSigDB; refs. 31, 32) to investigate the 297 genes that modulated response to either BYL719 or GDC0941 and found enrichment for multiple pathways involved in PI3K signaling, including mTOR, AKT (PKB), and insulin signaling (Fig. 1H). Multiple RTK signaling pathways, including the ERBB family (in particular EGFR and ERBB2) as well as FGFR and PDGF, were enriched among the hit genes that were significantly enriched or depleted from the genome-wide CRISPR screen (FDR < 0.3).

“Signaling by EGFR in cancer” was the most highly enriched pathway and is of particular interest in pancreatic cancer as the EGFR inhibitor erlotinib has shown some modest activity in patients with pancreatic cancer (43). Moreover, guides targeting genes associated with the internalization and degradation of activated ERBB family receptors were enriched in the drug-treated populations. Suppressor hit genes included AP2S1, AP2B1, and AP2M1, which encode subunits of the adaptor protein 2 (AP-2) complex and are involved in clathrin-dependent endocytosis of activated EGFR (44). We hypothesize that loss of the AP-2 complex would result in sustained EGFR signaling. PRKACA encodes a catalytic subunit of protein kinase A (PKA), which phosphorylates and inhibits EGFR (45), and facilitates its internalization and ubiquitination (46). Overall, loss of these genes may result in activation of EGFR, thereby promoting resistance to PI3K inhibition.

Penetrant synthetic lethal interactions, which demonstrate similar effects across diverse cellular models, may have greater therapeutic benefit as they could overcome the molecular heterogeneity that exists within tumors (47). Therefore, to prioritize penetrant synthetic lethal effects, we generated a minipool targeting 296 hit genes from the genome-wide screen, including those hits identified with either BYL719 or GDC0941, for a secondary validation screen. We tested this library in MIAPACA2 cells and in three additional KRAS-mutant pancreatic cancer cell lines. All cell lines chosen for the validation screen were of the QM subtype of PDAC as this represents the subtype with the poorest prognosis and therefore the most urgent clinical unmet need (48). We confirmed that Cas9 expression did not alter response to PI3K inhibition (Supplementary Fig. S3A). We selected concentrations of BYL719 and GDC0941 that inhibited PI3K signaling, but still permitted cell proliferation (Supplementary Fig. S3B). Cells were then transduced with the minipool library. After puromycin selection, cells were treated with DMSO, BYL719, or GDC0941 and passaged for approximately eight population doublings (Supplementary Fig. S4A). sgRNA abundance was determined as in the whole-genome screen. For each cell line, the abundance of nontargeting sgRNAs was not changed compared with the plasmid DNA, but sgRNAs targeting essential genes were depleted, indicating that transduction led to successful gene editing (Supplementary Fig. S4B).

STARS analysis was used to prioritize genes with multiple top-scoring sgRNAs that were either enriched or depleted from the drug-treated arms (Supplementary Tables S9–S24). To discover penetrant hits, genes were ranked according to their average STARS score across all four cell lines (Fig. 2A). Reassuringly, there was considerable overlap between the hit genes that could modulate sensitivity to BYL719 and GDC0941. MEMO1, UBE2H, MIOS, and YPEL5 were the top four hit genes that, when lost, sensitized to both BYL719 and GDC0941 across all four cell lines. Targeting of PTEN, TSC1, TSC2, FRYL, PDCD10, and NF2 were the top six hit genes that drove resistance to both PI3K inhibitors. Notably, sgRNAs targeting KEAP1 were enriched in the presence of PI3K inhibition suggesting resistance could also be driven by NFE2L2/NRF2-mediated activation of an antioxidant stress response pathway (49). As STARS only uses the top 10% of sgRNAs to rank genes, we also analyzed the minipool screen based on the average LFC of all sgRNAs for each cell line and then ranked each gene in the minipool based on the average LFC across all cell lines. Reassuringly, both analysis approaches showed agreement (Supplementary Fig. S4C). We also confirmed that sgRNAs targeting the top-ranking genes identified in the primary screen also showed significant enrichment or depletion in the MIAPACA2 cell line in the secondary screen (Supplementary Fig. S4D). This suggested good concordance between the primary and secondary screens in MIAPACA2 cells.

We focussed on two hits – MEMO1 and UBE2H, as they were both related to ERBB family signaling. Knockout of MEMO1 or UBE2H by CRISPR/Cas9 was confirmed (Fig. 2B) and enhanced the antiproliferative effect of BYL719 in PANC03.27 and MIAPACA2 (Fig. 2C). Both MEMO1 and UBE2H regulate signal transduction by the ERBB family and IGF1R (50, 51); therefore, we hypothesized that stimulation of RTKs with specific ligands could promote resistance to PI3K inhibition. First, by culturing cells in low serum (0.1% FBS) AKT, PRAS40 and S6 phosphorylation were all decreased (Supplementary Fig. S5A), suggesting removal of growth factors could dampen signaling, even in the setting of oncogenic KRAS. Furthermore, cotreatment with BYL719 caused a near-complete suppression of AKT, PRAS40, and S6 phosphorylation. Low serum reduced cell proliferation by approximately 50% relative to 10% serum as did treatment with BYL719 (Supplementary Fig. S5B). Low serum and BYL719 treatment modestly suppressed cell proliferation further but to a lesser degree compared with drug treatment in 10% serum. This may reflect either the reduced proliferation rate of the cells in low serum and/or a decrease in PI3K signaling under low serum conditions. Interestingly, the addition of EGF, heregulin HRG and insulin-like growth factor 1 (IGF1) significantly increased the GI₅₀ concentration for BYL719 (Fig. 2D). HRG conferred the greatest degree of resistance to BYL719, associated with sustained AKT and S6 phosphorylation in the presence of BYL719 (Fig. 2E). Notably, despite EGF strongly activating the EGFR receptor and causing the expected downregulation of EGFR expression (52), it was not as effective as HRG in driving resistance to BYL719. Taken together, these data suggest that the ERBB family can drive resistance to PI3K inhibition in PDAC cells.

To identify clinically relevant inhibitors of RTK signaling that synergized with PI3K inhibition, we used an established library of 485 FDA-approved drugs and tool compounds, alone and in combination with 10 μmol/L BYL719 or 1 μmol/L GDC0941. The library was screened in MIAPACA2 cells at a concentration of 800 nmol/L, a concentration empirically chosen for a balance between being sufficient to modulate the target in cells, but not so high as to induce off target effects. Nevertheless, some synergistic interactions may not be detected for those compounds that were used at a too high or too low concentration. The Bliss independence model (20) was used to calculate synergy for each drug combination (Supplementary Tables S25 and S26). To identify hits common to both PI3K inhibitors, the Bliss score for BYL719 was plotted against that for GDC0941 (Fig. 3A). The compounds were also ranked on the basis of their average Bliss score for both PI3K inhibitors (Fig. 3B). Notably, the ERBB family inhibitor pelitinib demonstrated greatest synergy with both PI3K inhibitors. Another ERBB family inhibitor, dacomitinib, also demonstrated synergy with both compounds. Multiple inhibitors of mTOR also ranked highly, including KU-0063794, rapamycin, ridaforolimus, everolimus, and WYE-354. KU-0063794 inhibits mTORC1 and mTORC2 kinase activity and, given that it drove greater synergy than mTORC1 allosteric inhibitors such as rapamycin, suggests that dual mTORC1/2 inhibitors may elicit greater synergy with PI3K inhibitors. The ERBB family inhibitor, pelitinib (53), and the mTORC1/2 kinase inhibitor, AZD2014 (54), were selected to validate the synergistic interaction between inhibition of PI3K and ERBB or mTOR signaling. The combination of BYL719 and pelitinib or AZD2014 synergistically inhibited proliferation of pancreatic cells in both short- (Fig. 3C) and long-term (Fig. 3D) assays. This highlighted the capacity of the ERRB family and the mTOR pathway to drive resistance to PI3K inhibition.

Using a spheroid coculture of MIAPACA2 cells with activated PSCs—thought to better-model tumor–stromal interactions and the 3D tumor environment in vivo compared with 2D culture on plastic (55, 56), the combination of BYL719 and pelitinib or AZD2014 robustly inhibited spheroid growth (Fig. 3E and F). Propidium iodide staining of spheroids after 4 days of treatment demonstrated a significant increase in dead or dying cells with the combination of BYL719 with either pelitinib or AZD2014 versus single agents (Fig. 4G). Overall, we have clearly demonstrated that combined inhibition of PI3Kα and either ERBB or mTOR is synergistic in multiple models of PDAC.

The CRISPR and drug combination screens suggested that under PI3Kα inhibition, pancreatic cancer cells depend on mTOR signaling for proliferation. Therefore, we hypothesized that inadequate suppression of mTORC1 signaling underlies the resistance of pancreatic cells to BYL719. Indeed, phosphorylation of S6 (Ser240/244, catalyzed by p70^S6K) was not suppressed by BYL719 treatment in resistant PDAC cells, but was seen in the sensitive breast cancer cell line, T47D, whereas phospho-AKT was suppressed in both the sensitive and insensitive lines (Fig. 4A). Moreover, while CRISPR knockout of p110α decreased phosphorylation of AKT (Ser473) and PRAS40 (Thr246), phosphorylation of S6 (Ser240/244) was maintained (Fig. 4A). Across a panel of 12 pancreatic cell lines, the inhibition of phospho-S6 (Ser240/244) achieved with 10 μmol/L BYL719 closely correlated with the effect on cell proliferation (Fig. 4B). This suggests that mTOR signaling is uncoupled from PI3K in pancreatic cell lines and that this limits response to PI3Kα inhibition. Hence, we suggest that inhibition of phospho-S6 (Ser240/244) is an important and independent predictor of response to BYL719 versus other more proximal markers of PI3K signaling.

Our results suggest that inhibition of PI3K alone does not inhibit proliferation and that combination with mTORC1 inhibition is required. However, clinical trials of allosteric mTOR inhibitors in pancreatic cancer have been unsuccessful likely due to loss of negative feedback on IRS1 (7, 16) and, as shown herein, pancreatic cancer cells are resistant to single-agent inhibition of mTORC1 in vitro (Supplementary Fig. S6A), despite suppression of S6 phosphorylation (Supplementary Fig. S6B). We propose that inhibition of both PI3K and mTORC1 signaling is essential to inhibit cell proliferation. In line with this, the mTORC1/2 kinase inhibitors dactolisib (BEZ235) and AZD8055 (a closely related analogue of AZD2014) displayed potent antiproliferative activity in pancreatic cancer cell lines (Supplementary Fig. S6C and S6D) and this was associated with inhibition of phospho-S6 (Ser240/244) and at approximately 10-fold higher concentrations, inhibition of phospho-AKT (Ser473), as expected by dual inhibition of mTORC1 and mTORC2 (Supplementary Fig. S6E and S6F).

We studied the effect of the combinations of BYL719 with pelitinib or AZD2014 on PI3K–mTOR signaling. BYL719 alone resulted in near-complete suppression of phospho-AKT (Ser473), but decreases in phospho-S6 (Ser240/244) were not sustained. (Fig. 4C; Supplementary Fig. S7A and S7B). Similarly, AZD2014 or pelitinib could not sustain inhibition of both PI3K-AKT and mTORC1 signaling for 72 hours. Only the combination of these agents with BYL719 was sufficient to durably inhibit signaling at both nodes (Fig. 4C; Supplementary Fig. S7A and S7B). Given that inhibition of ERBB signaling could also decrease MAPK pathway activity, we also assessed the effect of pelitinib alone and in combination with BYL719 on ERK1/2 phosphorylation. However, no robust inhibition was observed, suggesting that decreased MAPK pathway activity was not contributing to the antiproliferative activity of this combination (Supplementary Fig. S7B).

To seek clinical relevance for our findings, we investigated how expression of selected genes, implicated by both our CRISPR and drug screens, related to clinical outcomes in PDAC by interrogating publicly available TCGA “provisional data” in cBioPortal (57, 58). Ninety-one percent of patients in this dataset have KRAS-mutant tumors. Genetic alterations of the ERBB family and PI3K signaling axis were present in 40% of 149 cases (Fig. 5A) and associated with poor survival among patients with PDAC, with a significant decrease in median survival from 23 months to 16 months (Fig. 5B). Therefore, genetic alterations in the ERBB family and PI3K signaling pathways are common in patients with PDAC and may contribute to a poor clinical outcome.

Overcoming acquired resistance to targeted therapies is arguably the major challenge facing drug discovery for the treatment of cancer. As exemplified by our CRISPR and drug combination screens, mechanisms of resistance to PI3K inhibition in PDAC converge on signaling through mTORC1. Incomplete suppression of mTORC1 underlies intrinsic resistance to PI3K inhibition and correlates with drug response. This was also predictive of response to PI3Kα inhibition in cell lines and patient tumors in PI3K-dependent breast cancer (41). Taken together, this suggests that mTORC1 has utility as a biomarker of PI3K inhibition and loss of mTOR signaling, combined with PI3K inhibition, is necessary to inhibit tumor growth. Interestingly, a drug-modifier CRISPR screen with the KRAS inhibitor MRTX849 also identified mTOR depletion as an enhancer of drug activity and validated the combination of MRTX849 with everolimus and AZD2014. Consistent with our data, near-complete suppression of phospho-S6 was associated with an antiproliferative effect (6).

Our data suggest that alternative pathways may compensate for PI3K inhibition to reactivate mTORC1. PI3K signaling is regulated by growth factors, as removal of FBS is sufficient to inhibit signaling through AKT and PRAS40 in PDAC cells. EGF confers resistance to BYL719 in head and neck cancer (59) and IGF1 and neuregulin 1 (also known as heregulin, HRG) drive resistance to PI3K inhibition in PIK3CA-mutant breast cancer (41). IGF1 is of interest in pancreatic cancer as it is found at high levels in tumor stroma (60). The greatest protective effect was associated with reactivation of PI3K signaling by HRG and suggests that ligand-mediated ERBB family activation participates in driving resistance to PI3K inhibition in PDAC, even in the context of oncogenic KRAS.

Numerous regulators of RTK signaling were implicated in resistance to PI3K inhibition in both CRISPR screens. Of these, MEMO1 interacts with IGF1R and all four ERBB family members (50, 51) and mediates activation of MAPK and PI3K signaling (51). MEMO1 also interacts with IRS1 and prevents dephosphorylation and deactivation of IRS1 signaling (50). UBE2H is involved in insulin and PI3K signaling in skeletal muscle and cooperates with the E3-ubiquitin ligase, Mitsugumin 53 (MG53 or TRIM72), to ubiquitinate and downregulate IRS1, which is important for negative feedback regulation of IGF1 and insulin signaling and inhibition of skeletal myogenesis (61). Loss of these genes sensitized cells to PI3K inhibition, demonstrating that RTK signaling is a clear determinant of response to PI3K inhibition.

The combination drug screen suggested that targeting of mTOR or the ERBB family could circumvent resistance to PI3K inhibition. The ERBB family consists of four receptor tyrosine kinases, which are activated by ligand binding and regulate the RAS, MAPK, and PI3K pathways (62). However, KRAS is activated downstream of EGFR signaling, implying that EGFR may have little relevance in tumors driven by constitutively activated RAS signaling. In support of this, activating mutations in KRAS drive resistance to EGFR inhibitors in colorectal cancer (63, 64). Conversely, in PDAC clinical trials, addition of the EGFR inhibitor, erlotinib, to gemcitabine resulted in a modest survival benefit, suggesting that tumors may still partially rely on EGFR signaling (43). In line with this, we show that genetic alterations of ERBB and PI3K pathway members in PDAC patient tumors associates with poor survival and may provide potential patient selection criteria for this drug combination, which warrants further investigation as a novel therapeutic strategy in PDAC. The use of gene expression signatures, to classify pancreatic cancers into distinct subtypes that exhibit vulnerabilities to specific drugs, also has the potential to inform treatment decisions (48). For example, we focused on the QM subtype of pancreatic cancer, so this patient population would be a rational choice for preliminary investigations. Furthermore, around 50% of clinical pancreatic samples are EGFR positive and overexpression correlates with poor survival (43). Upregulation of EGFR occurs selectively in PanINs and early stages of PDAC in mice (65, 66), which implicates this receptor in tumor development. Moreover, mouse models show that Kras-driven tumorigenesis is dependent on Egfr, as genetic inactivation of Egfr blocks induction of PanINs and PDAC (65, 66). Similarly, studies of KRAS-mutant NSCLC show that ERBB family signaling amplifies the activity of mutant KRAS in in vivo models and loss of ERBB family signaling impairs tumor development (67, 68). Furthermore, studies conducted with the KRAS inhibitors AMG510 or MRTX849 have shown synergistic antiproliferative activity with ERBB family inhibitors (5, 6). Studies in pancreatic organoid models have shown synergy between either MEK or AKT inhibitors when combined with ERBB family inhibitors (but not with EGFR inhibition), underscoring the need to completely suppress ERBB signaling for activity (69). Pelitinib is a potent and irreversible EGFR inhibitor that also has activity against other ERBB family members, most notably HER2 (70, 71). Therefore, the synergistic activity is likely not due to inhibition of EGFR alone, but through more durable inhibition of all ERBB family receptors. These data clearly support a role for the ERBB family in mutant KRAS signaling and KRAS-driven tumorigenesis and cell proliferation across tumor types.

We have shown that pancreatic cancer cell lines are predominantly resistant to inhibition of PI3K via sustained mTOR signaling, despite effective inhibition of upstream PI3K signaling, indicating that alternative pathways can maintain mTORC1 activation and promote proliferation. Our genetic and pharmacologic data show that dual inhibition of PI3K and mTORC1 signaling achieves greater antiproliferative activity than targeting a single node. Notably, dual mTORC1/2 kinase inhibitors, such as AZD2014, achieve this at low nanomolar concentrations that are pharmacologically relevant, whereas higher concentrations of mTORC1 allosteric inhibitors are required to achieve similar effects. This suggests a potential benefit of inhibiting both mTORC1 and mTORC2. Furthermore, this can also be achieved by combined inhibition of PI3Kα and ERBB family signaling, indicating that the ERBB family is important for sustaining mTORC1 activity in the presence of PI3K inhibition, even in the context of mutant KRAS (Supplementary Fig. S8). Given the role of ERBB family activity in enhancing signaling through KRAS in NSCLC (67, 68), one may speculate that inhibition of mTOR, downstream of ERBB family inhibition, may result from attenuation of RTK-driven wild-type KRAS signaling (72). However, we did not detect robust inhibition of ERK phosphorylation in response to pelitinib (alone or with BYL719), suggesting this was not MAPK dependent. This study provides the basis for future translational work in xenograft and genetically engineered mouse models of pancreatic cancer, to determine the tolerability and efficacy of combined PI3K and mTOR kinase inhibitors or irreversible ERBB family inhibitors. Reassuringly, the combination of PI3K and pan-ERBB inhibitors has been tested in KRAS or PIK3CA-driven xenografts and genetically engineered mouse models confirming this therapeutic strategy is tolerated and efficacious in vivo (73–75). Our data suggest potential pharmacodynamic biomarkers to monitor drug response and guide dosing strategies. Positive results may renew interest in these classes of therapeutic agents for this challenging cancer type.

No potential conflicts of interest were disclosed.

Conception and design: C.K. Milton, S.R. Whittaker

Development of methodology: C.K. Milton, S.R. Whittaker

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): C.K. Milton, A.J. Self, D.E. Root, S.R. Whittaker

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): C.K. Milton, F. Piccioni, D.E. Root, S.R. Whittaker

Writing, review, and/or revision of the manuscript: C.K. Milton, P.A. Clarke, U. Banerji, S.R. Whittaker

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): A.J. Self, F. Piccioni, S.R. Whittaker

Study supervision: P.A. Clarke, U. Banerji, S.R. Whittaker

This work has been funded by The Institute of Cancer Research, Pancreatic Cancer UK and the Louis Nicholas Residuary Charitable Trust (to S.R. Whittaker). We thank Drs. Muge Sarper and Amine Sadok at The Institute of Cancer Research for advice on the spheroid, coculture assays and Drs Mark Stubbs and Rosemary Burke, also at The Institute of Cancer Research, for access to the compound library and assistance with ECHO dispensing.

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.