Phosphoproteomics Identifies PI3K Inhibitor–selective Adaptive Responses in Pancreatic Cancer Cell Therapy and Resistance

In pathophysiologic signaling, biochemical and biomechanical cues are integrated and regulated in the long term. Similarly, it is expected that inhibition of signaling pathways by targeted therapies toward one signal transduction enzyme also induces an adaptation of the entire signaling network.

Class I PI3Ks are crucial signal transduction enzymes. After acute stimulation, PI3K phosphorylates the lipid second messenger phosphatidylinositol 4,5-biphosphate into PI-3,4,5-triphosphate (PIP₃) at the plasma membrane, further activating the protein kinases Akt and mTOR, and regulating major cell biology events such as cell proliferation, cell survival, and protein synthesis. PI3K is one of the most altered pathways in cancers and presents four different isoforms encoded by four different genes (1, 2). Because of the regulation of fundamental cellular processes by PI3K/Akt/mTOR, this signaling axis is an excellent therapeutic target in cancer which is underscored by the number of molecules tested currently in clinical trials (3). Although class I PI3K isoform specificity is well described and accepted in physiology [for review: (1), examples: (4–7)], potential benefits of isoform-selective targeting in solid cancer were recently shown in breast cancers driven by oncogenic PI3Kα (8) but are still not yet approved in other genetic context promoting PI3K signaling (3). Besides, there are other effectors downstream PI3Ks than Akt/mTOR (9, 10). Such other downstream signaling routes are possibly contributing to the isoform specific in vivo role of mammalian PI3Ks. Although acquired cross-activation mechanisms between PI3K isoforms upon their unique and selective pharmacologic or genetic inhibition in the context of specific mutational landscapes (e.g., oncogenic PIK3CA, mutant PTEN) has been described previously (11–13), a specific large-scale cell signal adaptation to such treatment is unknown, in particular, in the context of nonmutated PI3Ks and/or Kras mutation. This genetic context is found in pancreatic ductal adenocarcinoma (PDAC). It is also unclear whether differing inhibition of each PI3K due to distinct PI3K inhibitor selectivity or due to different isoform expression could favor selective feedback mechanisms; the latter could be an unprecedented explanation of the therapeutic failure of PI3K inhibitors in unselected solid tumors.

New strategies are needed for the cure of patients with PDAC, due to dramatic lethality rate of this disease. PI3K signaling as assessed by Akt phosphorylation or by PI3K/Akt/mTOR gene signature is increased and associated with poor prognosis (14, 15). This increase in PI3K signaling is mostly due to Kras mutation engaging PI3Kα and possibly PI3Kγ (16–19). PI3K signaling downstream Kras is amplified by other signaling cues (20) and such amplification of Kras-PI3K coupling is critically involved in PDAC poor prognosis (21). Targeting PI3K is expected to have a clinical action in these patients (clinical trials pending as reviewed in refs. 22 and 23). However, prior knowledge of cancer cell adaptation to the inhibition of upstream class I PI3K signaling would be necessary to develop efficient anti-PI3K therapeutic strategy in pancreatic disease (22). The purpose of the study is to compare the pancreatic cancer cell phosphoproteome in response to inhibitors with different selectivity and off-target profiles, as an exploratory experiment to find first evidence of isoform-selective pathways and adaptive responses.

Therefore, we analyzed adaptive signals involved in response to inhibitors with distinct PI3K isoform selectivity in a human pancreatic cell line using a stable isotope labeling with amino acids in cell culture (SILAC)-based quantitative phosphoproteomics approach in a comprehensive way. By enrichment of the phosphoproteome followed by data mining, we demonstrate that, despite a common core pathway regulated by all inhibitors, the different PI3K inhibitors instruct specific, nonredundant signaling pathways linked to PI3K signaling. These data might help to better design therapeutic strategies in this dismal disease, including combination therapies against multiple isoforms of PI3K.

Reagents were purchased as follows: for in vitro assays, pan-PI3K and isoform-selective PI3K inhibitors from Axon Medchem; gemcitabine was a kind gift of hospital (IUCT-O, Toulouse, France); MTT was from Euromedex (4022). All PI3K inhibitors were resuspended in DMSO, corresponding at the vehicle condition. All products were resuspended according to the supplier's instructions.

Human pancreatic cell lines (Capan-1, BxPC-3, PANC-1, MIA PaCa-2) came from ATCC, human acute myeloid leukemia cell line (MOLM4) and murine pancreatic cancer cell lines (DT4994, DT6585, DT6606, DT8442, DT8661, R221, R259) were made in house or a kind gift from Dieter Saur (Klinikum rechts der Isar der TU München, Munich, Germany); for validation of genotype (16). Absence of Mycoplasma contamination was verified periodically by PCR and maintained in culture for a maximum of 15 passages after thawing. Capan-1, PANC-1, and MIA PaCa-2 cells were authenticated (STR method, Eurofins). Pancreatic cancer patient-derived organoids (PDO) were either derived from surgical resection (B25) or endoscopic fine-needle aspiration (B34) at Klinikum rechts der Isar der TU München, Munich, Germany. Patients were enrolled and consented in writing according to the Institutional Review Board approval project number 207/15 of the Technical University Munich (Munich, Germany). The studies were conducted in accordance with the Declaration of Helsinki. PDOs were characterized using whole-exome sequencing and RNA sequencing as described previously (24). Human normal and adenocarcinoma pancreatic samples (>30% tumoral cells) were selected at IUCT-O clinic, and collected according French and European legislation (CRB Biobank, France with following ethical authorization numbers BB-0033-00014, DC-2008-463, AC-2013-1955). KPC and KPC; p110γ^−/− mice were obtained and their genotype verified as described in ref. 25.

Human pancreatic cancer cell lines Capan-1 and BxPC-3 were cultured in RMPI1640 medium. PANC-1, MIA PaCa-2 and all murine pancreatic cancer cells were cultured in DMEM with 4.5 g of glucose (D6429, Sigma). All media were supplemented with 10% FBS (Eurobio), 1% glutamine (G7513, Sigma), and 1% antibiotics (penicillin/streptomycin, P0781, Sigma). Patient's organoids cell lines were cultured in special medium [Reichert composition (24)]. Specific methods to analyze cell proliferation/survival, mRNA and protein expression levels are described in Supplementary Materials and Methods.

Light amino acid–labeled and heavy amino acid–labeled Capan-1 cells (respectively called thereafter “light” and “heavy” cells) were cultured as described in Supplementary Materials and Methods. For each biological replicate, light lysates were mixed with the same heavy lysates at a 1:1 ratio for a total amount of 6 mg. Protein samples were reduced with 100 mmol/L DTT (Sigma, D9163) for 35 minutes at 57°C and then handled according to the FASP (filter aided sample preparation) digestion protocol (26) using Amicon Ultra-15 Centrifugal Filter device (10 kDa cutoff, MILLIPORE, UFC901096). Protocol for the enrichment with TiO₂ beads is based on Larsen and colleagues (27) and Jensen and colleagues (27). SILAC samples (TiO₂-enriched peptides) were resuspended with 2% acetonitrile, 0.05% TFA, and analyzed by nano-LC/MS-MS using an UltiMate 3000 system (Dionex) coupled to LTQ-Orbitrap Velos mass spectrometers (Thermo Fisher Scientific). All 40 mass spectrometry (MS) proteomic files have been deposited to ProteomeXchange Consortium with the dataset identifier as listed in Supplementary Table S1. For peptide identification, raw data files were processed in Proteome Discover 1.4.1.14 (Thermo Fisher Scientific) and searched against SwissProt human fasta database of Mascot (2014-06, sprot_20140428.fasta, 542782 sequences, high and medium confidence, Q value = 0.5–0.1). Peptides were further filtered using Mascot significance threshold S/N = 1.5 and a FDR <0.01 based on q-Value (Percolator). Phosphosite localization probabilities were calculated with phosphoRS 3.1 (maximum PTMs per peptide 10, maximum position isoforms 200). Phosphopeptides filtered with Proteome Discoverer 1.4.1.14 (see criteria in Supplementary Materials and Methods) were isolated from peptides. Only the ratios which were changed above and below the thresholds were processed for further analysis as described in Supplementary Materials and Methods.

All animal procedures were conducted in compliance with the Ethics Committee pursuant to European legislation translated into French Law as Décret 2013-118 dated February 1, 2013 (APAFIS 3601-2015121622062840).

Capan-1 cells were tested for their absence of Mycoplasma infection prior amplification and injection. A total of 3 × 10⁶ exponentially growing Capan-1 cells were subcutaneously in Nude/balbc Mice (Charles River, 9 weeks old, females). After 1 week implantation, we gavaged the mice 5 day a week with vehicle (0.5% methyl cellulose with 0.2% Tween-80) or with BYL-719/Alpelisib (25 mg/kg; MedChemExpress; ref. 28), IPI-549 (7.5 mg/kg; Biorbyt; ref. 29) alone or in combination. Toxicity parameters were assessed longitudinally with follow-up of mice weight, glycemia, blood cell counts [using Yumizen H500 hematology analyzer (HORIBA)]. The drug dosage with treatment 5 day a week lasted for 3 weeks, time at which the vehicle group had to be euthanized for ethical reasons. Tumor volume was measured with a caliper and calculated using the formula V = (4/3) × π × (length/2)2 × (width/2).

Statistically significant differences were performed with GraphPad Prism using the t tests (paired test): *, P < 0.05; **, P < 0.01; ***, P < 0.001. Nonsignificant (ns) if P > 0.05. For in vivo experiments, Mann–Whitney test was used: **, P < 0.01; P value is indicated for nearly significant values.

Phosphoproteomics data have been deposited in Pride/ProteomeXchange with the dataset identifier PDX008410 the 11/12/2017. PDO, plasmids for sh expression used in the article require an MTA.

In pancreatic cancer, PI3K signaling is associated with a poor prognosis (14, 15). The analysis of 11 pancreatic cancer samples compared with normal adjacent tissue showed increase Akt expression and a significant increase of S473 and T308 Akt phosphorylation by Western blot (WB) analysis (9/11 patients; Fig. 1A; Supplementary Fig. S1A). However, the observed increase of Akt phosphorylation was not always associated with a significant increase in the phosphorylation levels of canonical downstream targets, such as pPRAS40 or pS6K (Fig. 1A, right), emphasizing the heterogeneity of signaling targets downstream PI3Ks in a clinical setting.

The human pancreatic cancer cell line Capan-1 represents common genetic alterations found in PDAC including KRAS, TP53, and SMAD4 mutations (30), expresses all four PI3K isoforms (16, 31) and displays a proliferation rate similar to other human pancreatic cell lines (as verified in Supplementary Fig. S1B). Amongst the four PI3K isoforms responsible for the production of PIP₃ and Akt activation, we and others have identified PI3Kα and PI3Kγ to be involved in pancreatic carcinogenesis (16, 17, 25, 32, 33). Besides oncogenic Kras-driven activation of PI3K (20), stimulation by FBS induces the activation of receptor tyrosine kinases (RTK) and G protein–coupled receptors (GPCR) that increases pAkt level (7). Short-term (10 minutes) FBS stimulation induced a significant activation of class I PI3Ks as assessed by the phosphorylation of Akt and a known downstream effector, PRAS40 (Fig. 1B). We used a pan-PI3K/mTOR-targeting inhibitor that inhibits all PI3K isoforms at equal potency [LY-294002 from here on named pan-inh]. pan-inh completely abolished pAkt and pPRAS40 (Fig. 1B). Isoform-selective drugs targeting either PI3Kα [A66 (5 μmol/L), α-inh (34)], PI3Kβ [TGX-221 (0.5 μmol/L), β-inh (7, 35–37)], and PI3Kγ [AS-252424 (5 μmol/L), γ-inh (38)], but not of PI3Kδ inhibitor [IC-87114 (5 μmol/L), δ-inh], inhibited pS473Akt as well as pPRAS40 levels significantly after 10 minutes of FBS stimulation (Fig. 1B). PI3K inhibitors are still effective to inhibit pAkt when diluted in cell medium 24 hours prior to treatment in vitro demonstrating the stability of the compounds at long term (Supplementary Fig. S1C).

To explore the possibility of isoform-selective downstream pathways, we devised an exploratory strategy to identify adaptive response to PI3K inhibitors with varying selectivity and off-targets effects in a comprehensive fashion by defining phosphosite regulated signaling pathways in Capan-1 cell line (Fig. 1C). To quantify subtle differences between these conditions targeting the same enzymatic activity, we chose a SILAC-based quantitative approach combined to a phosphopeptide enrichment by TiO₂ allowing a robust S/T/Y phosphorylation quantification of thousands of proteins. We devised a super-SILAC approach (39, 40), in which we compared all the unlabeled treatment conditions to vehicle control, with the use of SILAC-labeled cells as a spike-in standard for accurate quantification of unlabeled samples (Fig. 1C; Supplementary Table S1). Incorporation of heavy isotopes was verified by MS after six passages (Supplementary Fig. S2A and S2B); and the metabolic labeling did not change the proliferation and morphologic properties of Capan-1 cell line (Supplementary Fig. S2C and S2D). All validating steps are detailed in Supplementary Materials and Methods and Supplementary Fig. S2. In all conditions combined, 3,600 heavy/light phosphopeptides were detected and quantified (Fig. 1C). Among these, 83% serine-sites (S), 16% threonine-sites (T), 1% tyrosine-sites (Y) were phosphorylated (Supplementary Fig. S2F). These percentages were unchanged upon PI3K inhibition (Supplementary Fig. S2G). Short-term (10 minutes) and long-term (24 hours) serum stimulation induced modifications (increased and decreased) of 557 phosphopeptides (28%) and 619 phosphopeptides (32%), respectively (Supplementary Fig. S2H and S2I).

Overall, levels of a known PI3K target PRAS40 were changed in similar manner albeit with a slightly lower dynamic range when quantified by MS-based proteomics analysis compared with WB results (Supplementary Fig. S2J vs. Fig. 1B).

We next identified phosphopeptides with significant altered levels in each condition in an unbiased fashion. PI3K inhibitors induced more phosphorylation level changes after 24 hours of treatment than after 10 minutes. Interestingly, γ-inh led to the most significant changes in numbers of significantly modified phosphopeptides compared with FBS as soon as 10 minutes of treatment, and global equivalent phosphoprotein level changes at 24 hours compared with α-inh (Fig. 1D). These strong phosphoprotein modulations by γ-inh and α-inh after 24 hours of treatment suggest signaling engagement of these two PI3K isoforms in PDAC cells, in addition to their involvement in pancreatic carcinogenesis (16, 17, 25, 32, 33).

The number of significantly increased phosphopeptides in γ-inh condition (and to slighter extent of pan-inh) was higher at short time compared with α-inh but decreased at 24 hours, possibly suggesting an early induction of feedback control with these two PI3K inhibitors. Reversely, α-inh treatment led to increased phosphopeptides at 24 hours. β-inh treatment led to a balanced increase/decrease of phosphopeptides at both times (Fig. 1D). Early upregulated signaling is expected to reduce the efficiency of PI3Kγ inhibitors.

We next analyzed dynamic changes in the phosphoproteome upon PI3K isoform-specific inhibition over time (10 minutes vs. 24 hours treatment) using global pathway analysis (i) to identify selectivity in regulating biological functions by each inhibitor, and (ii) visualize entire signal network rewiring upon PI3K inhibition pressure.

A principal component analysis (PCA) of the phosphoproteome at 10 minutes demonstrated that serum-stimulation alone, α-inh and β-inh conditions clustered together while pan-inh and γ-inh conditions separated from the cluster (Fig. 2A, left). After 24 hours of treatment, however, all inhibitor treatments separated from the FBS condition, highlighting the time necessary to induce significant PI3K inhibitor–selective changes in signaling networks upon PI3K inhibition, particularly for PI3K isoform-specific inhibition (Fig. 2A, right). Surprisingly, inhibition with LY (pan-inh) clustered closely together with the β-inh, while γ- and α-inh conditions led to distinct phosphopeptide modifications, in line with the known selective roles of these isoforms in PDAC (17, 31, 32).

Next, we identified inhibitor selectivity in regulated phosphoproteomes at both timepoints using hierarchical clustering (Supplementary Materials and Methods; Supplementary Tables S1–S3). By overlapping phosphopeptides with altered levels in each condition, we were able to identify shared (core) phosphopeptides, inhibitor-selective phosphopeptides indicated as either A66-selective (α-inh select), TGX-221–selective (β-inh select) or AS252424-selective (γ-inh select), as well as pan/mTOR-selective (pan-inh select) phosphopeptides (Fig. 2B and C; Supplementary Fig. S3). The finding that there is a common “core” of phosphopeptides regulated by all PI3K inhibitors was underscored by STRING analyses of inhibitor-selective and common phosphopeptides (Supplementary Fig. S3A and S3B; Supplementary Tables S1–S3). This core network presented a strong connectivity at 10 minutes of treatment.

Our data also demonstrate that each inhibitor induced distinct changes of the phosphorylation-regulated proteome already at 10 minutes of treatment (Fig. 2A and B), despite having similar effects on the phosphorylation of Akt and PRAS40 (Fig. 1B). Inhibitor-selective as well as core phosphopeptides showed distinct pathway enrichments (Reactome), cellular components and molecular functions (Gene Ontology) at the 10 minutes and 24 hours timepoints (Fig. 2C and D; Supplementary Fig. S3). At 10 minutes, common core downstream signaling includes 115 phosphoproteins implicated in mRNA splicing, chromatin regulation, TGFβ signaling as well as gene transcription and transcript processing pathways (ZE P value ≤ 0.05), such as EIF3G, EIF2S2, and EIF4G1 (Fig. 2C and D; Supplementary Table S2). Interestingly, at 24 hours, γ-inh–selective phosphopeptides were enriched in similar functions and cellular components as those changed by all PI3Ks/mTOR (core and pan-inh select), with targets presenting a strong protein–protein interactivity; those included CTNND1, JUP, and SRRM2 proteins (Fig. 2C and D; Supplementary Fig. S3F–S3J; Supplementary Table S2). Apart from the core pathways, α-inh–selective and β-inh–selective phosphopeptides were enriched in different pathways at both timepoints. Specifically, α-inh modulated phosphopeptides regulating Rho GTPases signaling, RTK and cytokine signaling (including RACGAP1, IRS2, STAT3), whereas β-inh regulated phosphopeptides directing mitotic control including targets such as RB1 (Fig. 2C and D; Supplementary S3D, S3E, S3G–S3J; Supplementary Table S2).

To estimate on-target and off-target effects of inhibitors, we used STRING connectivity analysis and showed that inhibitor selective targets were associated and connected with PI3K isoforms, except for TGX-221 treatment at 24 hours (Supplementary Fig. S3E; Supplementary Tables S1–S3). The proteins involved were either off-targets of TGX-221 (β-inh) or were unknown targets of PI3Kβ. These data show that PI3K inhibition with AS252424 (γ-inh) regulates similar processes to pan-inh and, at the same time, controls additional core networks exceeding the effect of pan-inh on signal network as shown by the number of nonselective and selective phosphopeptides modified by γ-inh treatment (Fig. 2B).

These findings suggest that inhibiting strongly PI3Kγ might be the most effective therapeutic strategy in this cell line to inhibit the core PI3K signaling to target essential growth and survival pathways in pancreatic cancer.

We next analyzed the expression levels of PI3K catalytic subunits that constitute the four class I isoforms (α, β, γ, and δ) in tumor cell–enriched human pancreatic cancer samples compared with normal adjacent pancreas by WB analysis. We observed that p110β levels were increased only in five patients, while p110α, p110γ, and p110δ protein levels were significantly increased (in all but 1 patient; Fig. 3A). To further verify these findings in a pure population of cancer cells, we analyzed PI3K isoforms protein and mRNA levels in human (n = 4) and murine (n = 7) cell lines (derived from genetically modified mouse models of PDAC; Fig. 3B,–E). The myeloid cell line, MOLM-14, which is known to express p110γ in high abundancy served as positive control.

In contrast to tumor cell–enriched human pancreatic cancer samples, p110γ was only found detected in one of four human pancreatic cancer cell lines and in two of seven murine pancreatic cancer cell lines by WB analysis (Fig. 3B and C). Similarly, on a transcript level, two of four human PDAC cell lines and three of seven murine PDAC cell lines showed increase in p110γ mRNA expression by qRT-PCR (Fig. 3D and E). These data indicate that p110γ is overexpressed in patient-derived tumor cell–enriched PDAC samples and to a lesser extent in PDAC cells in vitro, suggesting a stronger role for p110γ in vivo and in patients, compared with cell culture conditions.

We next confirmed that genetic inactivation of PI3Kγ using a full knockout approach almost completely abolished cancer formation from precancer lesions (PanIN) in a mutated KRAS and p53 background (Fig. 4A). Similarly, despite the low expression levels of p110γ in vitro, decreased expression of p110γ using a short hairpin RNA (shRNA) specific for p110γ reduced the clonogenicity of Capan-1 cell line (Fig. 4B). Of note, we did observe decreased PI3Kα expression upon p110γ knockdown by shRNA in one pool of cells, but that led to similar effects that the other shp110γ pool of cells.

We tested isoform-selective inhibitors on a panel of Human pancreatic cancer cells lines. γ-inh significantly decreased pAkt at 10 minutes in the four cell lines (Fig. 4C; Supplementary Fig. S4A). These experiments aimed to validate the pathway analysis on inhibitor-selective phosphoproteome, by comparing with experimental cellular outputs.

β-inh led to increased cell numbers in Capan-1 cells only (Fig. 4D). BxPC-3 cells (nonmutated KRAS cell line) was sensitive to all tested inhibitors (Fig. 4D). Cell numbers of all human pancreatic cancer cell lines were significantly decreased in time upon γ-inh or pan-inh as compared with vehicle, regardless p110γ mRNA level of expression, while α-inh was most efficient in BxPC-3 and Panc-1 cell lines (Fig. 4D). γ-inh was almost as effective or more effective (Capan-1) that pan-inh. This was confirmed with BrdU incorporation assay, cell-cycle analysis, and cleaved caspase-3 analysis (Supplementary Fig. S4B–S4D). Interestingly, only γ-inh treatment induced significant increase of DEVDase activity (Supplementary Fig. S4D), confirming the selective enrichment of the “Apoptotic cleavage of cellular proteins” signaling pathway by γ-inh (Fig. 2C, right).

We next analyzed the activation of PI3K canonical pathway at 24 hours. In Capan-1 cells, 24 hours inhibition with pan-inh or γ-inh led to a reactivation of p-Akt (Fig. 5A). α-inh treatment did not show such upregulation of pAkt, in the four tested cell lines (Fig. 5A). The combination of α-inh and γ-inh prevented this reactivation at 24 hours compared with γ-inh treatment. The reactivation of pAkt upon γ-inh or pan-inh treatment was observed in three and two of four cell lines, respectively (Fig. 5A; Supplementary Fig. S4A).

Interestingly, if we were able to identify selectively changed phosphopeptides in all conditions tested, pan-inh or γ-inh conditions also led to the identification of higher numbers of selective phosphoproteins (Fig. 2B), that are known to be interregulated (Supplementary Fig. S3B, S3C, and S3F), and leading similar pathway enrichments (Fig. 2C). This could also correspond to the induction of negative feedback loops. Indeed, kinase enrichment analysis showed that pan-inh and γ-inh treatment induced the phosphorylation of peptides that corresponds to RPS6KB2 (S6K2, mTORC1 downstream effector) kinase motif (P < 0.05) and to MAPK9 with lower statistical power (Supplementary Fig. S5). These bioinformatic data are line with the observed reinduction of pAkt upon pan-inh or γ-inh treatments (Fig. 5A).

Analysis of phosphopeptides regulated selectively by PI3Kα inhibitor at 24 hours in Fig. 2C and in Supplementary Fig. S3D showed that PI3Kα regulated additional functions and cellular components compared with panPI3K/mTOR or PI3Kγ-selective phosphopeptides. We thus tested the combination of α-inh and γ-inh on cell survival and showed that this combination led to a higher efficiency than each inhibitor alone or than PI3K/mTOR inhibitor in three of four cell lines (Supplementary Fig. S6). The only cell line (MIA Paca-2) where the combination was not more effective on cell survival was the one where we did not detect a feedback loop on pAkt upon PI3K inhibition. Combination of γ-inh with other isoform PI3Kβ or PI3Kδ inhibitors appeared to be less synergistic (Supplementary Fig. S6). This corroborates the phosphoproteomic analysis that identifies feedback mechanisms on S6K upon PI3Kγ inhibition.

We next asked the question whether the phosphopeptides selectively regulated at long-term by either PI3Kα or PI3Kγ inhibitors correlate with the effect of each PI3K inhibitor. The changes in ratios of pGIGYF2 (T382) correlated with PI3K action on pancreatic cell survival/proliferation (Supplementary Fig. S7A; WO2019101871-A1). Interestingly, changes of phosphorylation levels of known targets of PI3K as assessed by WB (Supplementary Fig. S4A) did not correlate with the cellular action in the four pancreatic cancer cell lines (Supplementary Fig. S7B).

Finally, we aimed to test this synergy in a clinically relevant setting, in two pancreatic cancer PDOs (B25 and B34) that displayed representative genetic alterations and p110s expression (Fig. 5B and C). Here, we observed that the equimolar combination of PI3Kα-inh and PI3Kγ-inh led to an increased sensitivity to PI3K inhibition as assessed with the metabolism measurements of organoids (Fig. 5D and E). In cell lines and PDOs tested, combined treatment had additive (2/6) or synergistic effect (4/6 including the two PDOs) at high concentration (5 μmol/L; Fig. 5F; Supplementary Fig. S8A). At lower concentration (500 nmol/L), combination was antagonistic for the cell lines except for the nonmutant KRAS cell line BxPC-3 and the PDO B34.

For the in vivo proof of concept, we used alone or in combination two isoform-selective inhibitors that are approved for cancer treatment (8) tested in PDAC (41) or in phase II of clinical trial (https://clinicaltrials.gov/ct2/show/NCT03961698), respectively, the α-selective inhibitor BYL-719/Alpelisib (25 mg/kg) and the γ-selective inhibitor IPI-549 (7.5 mg/kg). We confirmed with these inhibitors (Supplementary Fig. S8B and S8C) the results obtained with A66 and AS-252424 in Capan-1 as shown in Fig. 4C and D and Fig. 5A. The combination of therapeutic agents that inhibits potently PI3Kα and PI3Kγ was well tolerated and statistically more efficient that the treatment of each inhibitor alone (Fig. 5G; Supplementary Fig. S8D).

Overall, we observed a good correlation between the global phosphorylation-regulated pathway analysis and measured cellular and tumoral effects upon PI3K inhibitor–selective pressure, providing the first evidence of isoform-selective downstream pathways. These data also suggest that complete inhibition of PI3K with strong efficiency on PI3Kα and PI3Kγ is the most effective strategy for patients with pancreatic cancer (WO2019073031-A1). This specific combination suppresses PI3K/Akt pathway in long-term manner, decreases efficiently cell survival and prevents tumor cell adaptation to PI3K signal blockage.

Whether critical signal nodes can be circumvented is a fundamental question in tumoral biology and therapy resistance. Using a phosphoproteomics screen, we demonstrate that each PI3K inhibitor with distinct isoform selectivity displays differential effects on the phosphoproteome of pancreatic cancer cells. Reciprocal class I PI3K signaling is the exclusive product of inhibition-imposed pressure. Accordingly, a selective and strong inhibition of PI3K isoforms is able to induce selective pathway rewiring which could facilitate selective resistance (Fig. 2D).

Preclinical data in pancreatic cancer have indicated for a long time that PI3K could be a suitable target in patients with PDAC [initial study using Wortmannin in xenografts of human pancreatic cancer cell lines (42)]. There is now an urgent need to prevent feedback loops involved in pancreatic cancer progression upon pan-PI3K treatment. One strategy consists of identifying tumors which are more likely to respond. Indeed, it was recently shown that p27kip1, encoding a cyclin-dependent kinase inhibitor, facilitated NVP-BEZ235 (PI3K/mTOR inhibitor) sensitivity in a gene dose-dependent fashion and knockdown of p27kip1 decreased NVP-BEZ235 response in murine PDAC cell lines (43).

In pancreatic cancer cells, dual PI3K-mTOR inhibition induces rapid overactivation of MAPK pathway (44) whereas the treatment with PI3K and MAPK inhibitors were more efficient in preclinical models when used in combination than alone (44–46). Our kinase prediction data at 24 hours seem to indicate that both pan-PI3K and PI3Kγ-selective inhibition appear to activate MAPK9, possibly activating c-Jun pathway. Interestingly, these data are in line with our previous work where PI3K-driven NFκB activation negatively controls JNK activation (47). In a future work, those combinative strategies could be tested with PI3K isoform inhibitors.

We propose that, further to MAPK signaling (44–46), adaptive response to PI3K inhibition also unleash compensatory inter-isoform selective signaling that is dependent on the isoform selectivity of the inhibitor; this could be circumvented by well-balanced isoforms-specific PI3K inhibitors. Given our results obtained in vitro with the first generation of isoform-selective PI3K inhibitors, we are convinced that the clinically approved Alpelisib/BYL-719 (8) combined with IPI-549 (25) that we tested in a xenograft model should be further evaluated in more complex preclinical models (e.g., KPC mice) as well as in PDAC clinical studies.

Our group and others have mostly described: either (i) immediate compensations/redundancy between isoforms that are activated through similar mechanisms [e.g., PI3Kα and PI3Kδ downstream RTK (48); or PI3Kγ and PI3Kβ downstream GPCR (7)], or (ii) delayed (in 1–2 week time) compensation/redundancy between the two ubiquitously expressed PI3K, PI3Kα, and PI3Kβ, via genetic induction of their selective of mode activation (e.g., overexpression of RTK, mutation of PTEN; refs. 11, 12, 49). We identify here a possible novel mode of resistance, which is based on rapid rewiring network. Next study will need to confirm that it is isoform selective and not inhibitor selective as well as dissect the possible mechanisms of signal rewiring.

Given the fact that PI3Kγ and PI3Kβ isoforms are both downstream of GPCRs (7), we were surprised to observe that the inactivation of these two isoforms did not lead to similar phosphoproteomic profiles, nor to similar alteration of cellular functions. One explanation of the poor effect of TGX-221 on these cells despite the observed reduced p-Akt levels could be the concomitant increased number of significantly increased phosphopeptides.

Of note, based on our results, isoform cross-compensation (in a context of nonmutated PI3K or of no loss of PTEN and of KRAS-mutant or not) appears to involve PI3Kα and the lowly expressed PI3Kγ. Because low level of expression of p110s could be difficult to detect, our data show the importance of determining the level of activation of each class I PI3K possibly with isoform selective gene signatures representative of their activity (50). It is also surprising that despite PI3Kγ having low levels of expression, its selective inhibition in serum-deprived condition was effective in all tested cell lines as others reported (32, 51). Importantly, the level of an identified PI3Kγ-selective phosphopeptide (phopho-GIGYG2 T382) was correlated with PI3K sensitivity. We hence propose that the level of phosphorylation of this selective target could be tested as a predictive marker of sensitivity, attesting the induction of selective compensation and efficiency of combinatory treatment with PI3Kα and PI3Kγ inhibitors.

Specific engagement of either PI3Kα or PI3Kγ downstream different types of Kras mutations was described recently (18, 19). It is also interesting to note that the synergistic effect of the tested combination was found in the mutated Kras cell lines, and that the PDO with increased frequency of Kras allele mutations (B34) displayed an increased combination index. These intriguing data prompt the better delineation of PI3Kα and PI3Kγ possible cooperation in various Kras-mutant contexts.

So far, only multicombinatorial therapies displayed positive clinical outcome for patients with PDAC (22). Isoform-specific drugs are expected to induce fewer secondary effects (3) and could thus be included in these multidrug combinatorial therapies. As we show that selective inhibitors at high doses induce selective feedback and resistance mechanism, these can be counteracted with the increasing arsenal of PI3K isoform-selective agents that are available for clinical use, in particular with Alpelisib/BYL-719 and IPI-549 (8, 25). In sum, results shown here highlight that defining pharmacologic profiles that are well-balanced toward each class I PI3K isoforms is key to therapeutic success.

C. Cintas reports grants from Ligue Contre le Cancer during the conduct of the study, has a patent for EP17306391 issued, and a patent for EP17306625 issued. T. Douche reports a patent for WO-2019101871-A1 issued, to INSERM/UPSIII/CNRS. M.-P. Bousquet reports a patent for WO2019/101871Al issued. C. Cayron reports grants from Ligue Nationale Contre le Cancer during the conduct of the study; grants from Ligue Nationale Contre le Cancer, Ligue Régionale Contre le Cancer, and grants and personal fees from FONROGA outside the submitted work. F. Ramos-Delgado reports grants from MSCA-ITN-PhD during the conduct of the study and grants from La Ligue Nationale Contre le Cancer outside the submitted work. O. Burlet-Schiltz reports a patent for WO2019/101871Al issued to Inserm. E. Hirsch reports personal fees from Kither Biotech during the conduct of the study. B. Thibault reports grants from Fondation RITC, Université de Toulouse (Emergence), Ligue Nationale Contre le Cancer, Fondation de France, MSCA-ITN-PhD, and COST PanGenEU during the conduct of the study. M. Reichert reports honoraria from Celgene, Roche, and Falk e.V. J. Guillermet-Guibert reports grants from RITC, Université de Toulouse, Ligue Nationale Contre le Cancer, Fondation pour la Recherche Médicale (FRM), Fondation FONROGA, Fondation de France, MSCA-ITN-PhD, COST PanGenEU, and other support from Forcheur Jean-Marie Lehn during the conduct of the study; in addition, J. Guillermet-Guibert has a patent for WO2019101871-A1 issued and a patent for WO2019073031-A1 issued. No disclosures were reported by the other authors.

C. Cintas: Formal analysis, investigation, visualization, methodology, writing–review and editing. T. Douche: Formal analysis, investigation, methodology, writing–review and editing. Z. Dantes: Investigation, writing–review and editing. E. Mouton-Barbosa: Formal analysis, methodology, writing–review and editing. M.-P. Bousquet: Formal analysis, methodology, writing–review and editing. C. Cayron: Investigation, writing–review and editing. N. Therville: Software, investigation, methodology, writing–review and editing. F. Pont: Software, investigation, methodology, writing–review and editing. F. Ramos-Delgado: Investigation, writing–review and editing. C. Guyon: Investigation, methodology, writing–review and editing. B. Garmy-Susini: Methodology, writing–review and editing. P. Cappello: Investigation, writing–review and editing. O. Burlet-Schiltz: Supervision, writing–review and editing. E. Hirsch: Resources, writing–review and editing. A. Gomez-Brouchet: Resources, writing–review and editing. B. Thibault: Supervision, investigation, methodology, writing–review and editing. M. Reichert: Conceptualization, resources, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. J. Guillermet-Guibert: Conceptualization, resources, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing.

We thank members of SigDYN group for constructive discussions, as well as members of CRCT core platforms, vectorology, cytometry and imaging, and B. Couderc for p110s-shRNA plasmids.

SigDYN group is a member of Labex TouCAN, ANR Programme D'excellence on resistance to therapies in cancer. The laboratory of J. Guillermet-Guibert for this project was funded by RITC, Université de Toulouse (Emergeance), Ligue Nationale Contre le Cancer (salary to C. Cintas and C. Cayron), Fondation pour la Recherche Médicale FRM, Inca (salary to T. Douche), Fondation FONROGA (C. Cayron), Fondation de France (salary to B. Thibault), MSCA-ITN-PhD (salary to F. Ramos-Delgado), COST PanGenEU and Université Paul Sabatier for French-German student exchange (to C. Cintas). M. Reichert is supported by the German Research Foundation (Deutsche Forschungsgemeinschaft, SFB1321 Modeling and Targeting Pancreatic Cancer Project-ID 329628492 and RE 3723/4-1). M. Reichert is supported by the German Cancer Aid Foundation (Max Eder Program, Deutsche Krebshilfe 111273). The laboratory of O. Burlet-Schiltz was supported in part by the Région Midi-Pyrénées, European funds (Fonds Europeéns de Développement Régional, FEDER), Toulouse Métropole, and by the French Ministry of Research with the Investissement d'Avenir Infrastructures Nationales en Biologie et Santé program (ProFI, Proteomics French Infrastructure project, ANR-10-INBS-08). J. Guillermet-Guibert and M. Reichert were awarded a Forcheur Jean-Marie Lehn Award on this project.

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.