Sustained Aurora Kinase B Expression Confers Resistance to PI3K Inhibition in Head and Neck Squamous Cell Carcinoma

Head and neck squamous cell carcinoma (HNSCC) is common, lethal, and disfiguring cancer with no biomarker-selected, molecularly targeted therapies available. The most recent genomic information available for this disease has not been translated into clinical care largely because the genomic landscape is dominated by tumor suppressors, including NOTCH1, which is mutated in about 18% of HNSCCs (1–4). The pattern of truncating and missense NOTCH1 mutations (NOTCH1^MUT) and supporting laboratory data demonstrate its role as a tumor suppressor in HNSCC (5).

To target genomic alterations in HNSCC, we recently assessed the degree to which responses to PI3K/mTOR pathway inhibitors were associated with gene mutations, mRNA and protein expression in 59 HNSCC cell lines. Remarkably, HNSCC cell lines harboring NOTCH1^MUT were significantly more sensitive to six drugs targeting PI3K or PI3K/mTOR than wild-type NOTCH1 (NOTCH1^WT) cell lines (6). In addition, NOTCH1^MUT HNSCC cells treated with PI3K inhibitors underwent significant apoptosis in vitro and in vivo. In contrast, HNSCC cells with PIK3CA mutations exhibited growth arrest, but not cell death, when treated with PI3K inhibitors (7). A clinical trial testing a dual PI3K/mTOR inhibitor in patients with NOTCH1^MUT HNSCC with recurrent or metastatic disease showed modest single-agent clinical activity (NCT03740100; ref. 8).

Modest clinical responses and acquired resistance (AR) are the leading causes of failure for otherwise promising and well-tolerated molecular targeted therapies, highlighting the importance of understanding molecular mechanisms of drug resistance. Furthermore, mechanisms of AR may overlap with innate resistance in patients with suboptimal initial responses. In this regard, potential mechanisms of resistance to PI3K inhibitors in NOTCH1^WT HNSCC remain unknown, and this represents a major gap in knowledge. We hypothesized that differentially regulated pathways following PI3K inhibition in sensitive and resistant HNSCC cell lines mediate resistance. In the current study, we determined the expression of over 300 proteins and phosphoproteins in both sensitive NOTCH1^MUT and resistant NOTCH1^WT HNSCC cell lines after PI3K inhibition to identify the key downstream pathways involved in PI3K inhibitor–induced apoptosis. We then examined whether NOTCH1^WT HNSCC could be sensitized to PI3K inhibition through simultaneously targeting of key downstream signaling pathways that mediate apoptosis in drug-sensitive NOTCH1^MUT HNSCC. We further tested whether these combination therapies could maximize killing of drug-naïve NOTCH1^MUT HNSCC and overcome AR that develops after prolonged single-agent PI3K inhibitor treatment.

Aurora kinase B (Aurora B, AURKB) was identified as a key effector molecule downstream of PI3K, which was downregulated in drug-naïve NOTCH1^MUT HNSCC compared with drug-resistant derivatives. Co targeting Aurora B together with PI3K enhanced killing of drug-naïve tumors, including NOTCH1^WT HNSCC, and could reverse AR in NOTCH1^MUT HNSCC. Furthermore, we demonstrated a link between downregulation of Aurora B and PDK1, another important mediator of PI3K inhibition–induced apoptosis we previously identified (6).

To the best of our knowledge, this is the first study to identify Aurora B as a mechanism of resistance to PI3K inhibition in any cancer type. Because NOTCH1 loss-of-function mutations are common in other squamous cell carcinomas, including those of the skin (9), esophagus (10, 11), and lung (12), our findings may be broadly applicable to many patients.

A panel of 56 HNSCC were obtained and maintained in their respective growth media as previously described (13–15). HNSCC cell lines used extensively in this study—HN31, UMSCC22A, PCI-15B, FaDu, MDA1386TU—were a kind gift from Dr. Jeffrey Myers (MD Anderson Cancer Center), whereas HEK293 cell lines were purchased from ATCC. These cell lines were cultured in high glucose DMEM supplemented with 10% FBS, 1% penicillin and streptomycin and 2 mmol/L l-glutamine and maintained in a humidified incubator with 5% CO₂. All cell lines were genotyped by short tandem repeat analysis and were Mycoplasma-free at the time of testing according to the Mycoplasma Detection Kit (LT07–705, Lonza).

We measured protein levels using reverse phase protein array (RPPA) with a panel of 304 antibodies (Supplementary Table S1; ref. 16) as previously described and performed immunoblotting as explained briefly below. For RPPA data processing, scanned images after hybridization and labeling were quantified using commercial software, MicroVigene (http://www.vigenetech.com/MicroVigene.htm) and SuperCurve (17). Our method iteratively fit joint logistic models to all data on a slide and returned summary estimates of log₂ protein expression values for each sample. The expression values were normalized across slides using a median centering approach to adjust for variability in sample loading, inducing consistent differences affecting all arrays in a set.

Doxycycline-inducible lentiviral vector for FOXM1b expression in mammalian cells (pCW57.1-FOXM1b) was obtained from Addgene. Hemagglutinin (HA)-tagged Aurora B plasmid was a kind gift from Dr. Ramon Parsons (Icahn School of Medicine, Mt. Sinai, NY). The drugs used in the current study were purchased form Selleck Chemicals and MedChem Express (Supplementary Table S2).

All animal studies were approved by the Institutional Animal Care and Use Committee at The University of Texas MD Anderson Cancer Center and are detailed in the Supplementary Methods. PASS 13 power analysis and sample size software (2014; NCSS, LLC) was used for the power/sample-size analysis, and investigators carried out these studies unblinded. Subcutaneous xenograft models were generated by injecting NOTCH1^MUT cells into athymic nude mice. Briefly, HN31 (0.75 × 10⁶ cells) or UMSCC22A (3 × 10⁶ cells) were injected subcutaneously on the right flank of the mice. Once the tumor volume reached ≥ 60 mm³, mice were randomized in their respective treatment groups. Mice injected with UMSCC22A cells were treated intraperitoneally with 10 mg/kg copanlisib (BAY806946) 3 times per week. Mice bearing HN31 tumors received 14 mg/kg copanlisib (BAY841236) intraperitoneally, 30 mg/kg alisertib by oral gavage, 50 mg/kg barasertib intraperitoneally, or a combination of copanlisib with alisertib or barasertib for 5 days per week for 21 days.

Apoptosis assays were performed as described previously (6). Briefly, TUNEL staining was carried out with either APO-BRDU (556405, BD Biosciences) or APO-DIRECT kit (556381, BD Biosciences), and Annexin V/propidium iodide (PI) staining was performed with a FITC-Annexin V apoptosis detection kit (556547, BD Biosciences) or PE-Annexin V apoptosis kit (559763, BD Biosciences) at indicated times. Samples were processed at the MD Anderson Flow Cytometry and Cell Imaging Core laboratory with a 3-laser, 10-color Gallios Flow Cytometer (Beckman Coulter) and analyzed using Kaluza software (Beckman Coulter).

Cells were lysed with ice-cold 1X RIPA buffer supplemented with protease and phosphatase inhibitors, and the lysates were centrifuged at 20,000 g for 10 minutes at 4°C. Lysates containing equal amounts of protein were resolved using SDS-PAGE, transferred to polyvinylidene difluoride membranes, and incubated with primary antibody overnight. Protein expression was detected using a horseradish peroxidase–conjugated secondary antibody (Bio-Rad) and electrochemiluminescence reagent (32106, ThermoFisher Scientific; or 1705062, Bio-Rad). The primary antibodies used in the study are listed in Supplementary Table S3.

Log₂ FPKMUQ values were calculated, and replicate samples were grouped by cell line and treatment. Low-expression genes were removed by filtering out any gene whose maximum group average across all treatments and cell lines was < 2 (log₂ space). Differences in gene expression between drug-treated and vehicle-treated values for each cell line were examined for statistical significance by performing multiple t tests and applying the Benjamini–Hochberg correction to control the FDR, and differences with an FDR < 0.1 were considered significant. Genes with significant differences were dichotomized into those upregulated (i.e., increased) or downregulated (i.e., decreased) following treatment with copanlisib to identify disjoint and common differentially expressed genes across cell lines.

As described previously (6, 16), cells were treated with DMSO or inhibitors as indicated at seven different concentrations (0–2 μmol/L) for 72 hours. Luminescence was measured using CellTiter-Glo (G7570, Promega) according to the manufacturer's instructions. For the combination screening in 56 HNSCC cell lines, the drexplorer R package with a best-fit dose response model was used to calculate inhibitory concentration values (18). The combination index (CI) values were calculated using the Chou–Talalay method (19–21) using Calcusyn software (Biosoft). CI values < 1 are considered synergistic.

Colony formation assay was performed in triplicates as previously described (6). Briefly, cells were seeded in 6-well plates and treated with DMSO or the indicated drugs. After 48 hours of treatment, media was replaced with drug-free media and cells were cultured for 14 to 21 days. The cell colonies were fixed and stained with crystal violet and analyzed using ImageJ software (NIH).

CellROX Green Flow Cytometry Assay Kit (C10492, ThermoFisher Scientific) was used for the detection of reactive oxygen species in live cells according to the manufacturer's instructions.

Cells expressing inducible FOXM1b were generated using lentiviral transduction and selected with 0.5 μg/mL puromycin. FOXM1b expression was induced with 100 ng/mL doxycycline. Cells constitutively expressing HA-tagged Aurora B, Aurora A, and AKT1 were generated using lentiviral particles obtained from the Functional Genomics Core laboratory at MD Anderson. Cells overexpressing HA-tagged Aurora B were selected with 2 μg/mL puromycin and cells with Aurora A and AKT1 overexpression were selected with 10 μg/mL blasticidin. For gene knockdown in HNSCC cells, siRNA was transfected with Lipofectamine RNAiMAX (13778–075, ThermoFisher Scientific) according to the manufacturer's instructions. Two specific siRNAs directed against AURKB and AKT1 were obtained from Horizon Discovery [AurB (1) - L-003326–00–0005; AKT1 (1) -L-003000–00–0005, siRNA SMARTpools), Invitrogen (AurB (2) – 4390824) and Santa Cruz Biotechnology (AKT1 (2) – sc-44198].

RPPA data analyses were performed using R (version 2.6.3). The two experimental factors in the RPPA data were phenotypes (NOTCH1^WT/NOTCH1^MUT cell lines) and PI3K treatment (vehicle/omipalisib; Fig. 1A). The limma package in Bioconductor (https://bioconductor.org/packages/release/bioc/html/limma.html) was used to compare changes in NOTCH1^WT and NOTCH1^MUT cell lines after treatment, and changes between the two (to identify differentially expressed proteins for treatment between phenotypes), which was the interaction term of the linear model. The Benjamini–Hochberg method (22) was applied to the resulting P values to control FDR. All statistical analyses were performed using R version 4.0.1 (R Core Team, 2020) or GraphPad Prism 7. An unpaired t test was used to compare the mean of two different groups when the distribution of the population was normal. One-way and two-way analysis of variance were used to compare the means of three or more groups under the assumption of normal distribution, and the Dunnett test or Tukey honestly significant difference test was applied for multiple comparisons. Kruskal–Wallis rank sum test was used to compare the means of more than two groups that didn't follow normal distribution. The Wilcoxon rank sum tests with Benjamini–Hochberg method were used to compare the pairwise groups after Kruskal–Wallis. All P values were two-tailed and for all analyses, P ≤ 0.05 was considered statistically significant, unless otherwise specified.

The data generated in this study are available within the article and its Supplementary Data files. Derived data supporting the findings of this study are available from the corresponding author upon request.

To investigate potential mechanisms mediating resistance to PI3K inhibition, we measured the levels of 304 proteins and phosphoproteins using RPPA analysis in three resistant NOTCH1^WT (FaDu, MDA1386TU, SCC25) and three sensitive NOTCH1^MUT (HN31, HN4, UMSCC25) cell lines after 4 and 24 hours of treatment with 50 nmol/L omipalisib (Fig. 1A), which is below its peak plasma concentration of 90 nmol/L (23). Omipalisib is a potent dual PI3K/mTOR inhibitor that was in clinical development at the time RPPA was performed. Many proteins were affected by PI3K/mTOR inhibition in both NOTCH1^MUT and NOTCH1^WT cell lines, including the expected changes in the PI3K/AKT/mTOR pathway, confirming appropriate and durable pathway inhibition (Supplementary Fig. S1A).

After 24 hours of treatment, 16 proteins were differentially expressed as a result of omipalisib treatment and NOTCH1^MUT status (P_interaction < 0.05 for treatment effect for NOTCH1^WT and NOTCH1^MUT) at FDR of 0.01, and 50 proteins were differentially expressed at FDR of 0.05 (Supplementary Fig. S1B and S2A). Differentially regulated proteins included expected markers of apoptosis (Mcl1, MDM2) and proliferation (p16, cyclin E1, PCNA, CDK1, Rb) in addition to Aurora B, forkhead box M1 (FOXM1), and several proteins involved in metabolism (glutamate D1, G6PD, ACC1). After 4 hours of treatment, 10 proteins were differentially regulated at an FDR of 0.05 (Supplementary Fig. S1C), but we did not study these because the magnitude of changes in these proteins was small and similar in NOTCH1^WT and NOTCH1^MUT cells.

For further study, we focused on proteins that were significantly altered after omipalisib treatment only in the NOTCH1^MUT cell lines and proteins that were regulated in opposite directions in the NOTCH1^WT and NOTCH1^MUT cell lines (Fig. 1B). We used immunoblotting to validate the RPPA findings in two NOTCH1^WT (FaDu, MDA1386TU) and three NOTCH1^MUT (HN31, UMSCC22A, PCI-15B) cell lines and confirmed consistent differences in the regulation of Aurora B and FOXM1 based on NOTCH1^MUT status following treatment with 50 nmol/L omipalisib for 24 hours (Fig. 1C and D). We also examined Aurora kinase A (Aurora A, AURKA), although it was not included in the RPPA, because Aurora A is known to regulate Aurora B (24). We detected similar changes in Aurora A protein levels in all NOTCH1^MUT HNSCC cell lines (Fig. 1C and D).

A recent study demonstrating destabilization of Aurora B levels through AKT degradation in PI3K/PTEN pathway mutant and wild-type KRAS/BRAF cancers (25) prompted us to investigate the total levels of AKT as well. We identified a remarkable decrease in total AKT levels across all NOTCH1^MUT HNSCC cell lines upon PI3K/mTOR inhibition (Fig. 1C and D). Consistent with our previous findings, we found both a significant downregulation in PDK1 levels (Fig. 1C and D) and evidence of apoptosis in NOTCH1^MUT lines treated with omipalisib (Fig. 1C; ref. 6). Moreover, RPPA showed significant changes in the levels of AKT and phospho-PDK1 selectively in NOTCH1^MUT HNSCC cell lines, although the P_interaction value was > 0.05 for total AKT (Supplementary Fig. S2B). In contrast, several proteins identified by RPPA were differentially regulated in some, but not all, of the cell lines based on NOTCH1^MUT status according to immunoblotting (Supplementary Fig. S1D).

To determine if these expression changes are specific to PI3K and confirm the robustness of our findings, we treated NOTCH1^WT and NOTCH1^MUT HNSCC cell lines in parallel with omipalisib or copanlisib, which is a potent, well-tolerated (26) pan-PI3K and FDA–approved drug. Both omipalisib and copanlisib reduced levels of Aurora A/B, FOXM1, PDK1, and total AKT to a much greater extent in NOTCH1^MUT than in NOTCH1^WT cell lines (Fig. 1E and F). In addition, copanlisib caused a significant increase in apoptosis in NOTCH1^MUT cells (Fig. 1G). Furthermore, we assessed the antitumor efficacy of copanlisib in a subcutaneous model of NOTCH1^MUT HNSCC (UMSCC22A) and found significantly lower tumor volumes compared with the vehicle-treated mice (Fig. 1H). These findings confirm that both pan-PI3K and dual PI3K/mTOR inhibitors cause apoptosis in NOTCH1^MUT HNSCC cells in vitro (6) and show that pan-PI3K inhibition is effective in vivo. For subsequent mechanistic experiments, we chose to continue with the more specific and clinically relevant drug, copanlisib.

We determined the dynamics of the differentially altered proteins through time-course analyses. Although the PI3K pathway was inhibited as early as 4 hours after drug treatment in all cell lines, protein levels of Aurora A/B, total AKT, FOXM1, and PDK1 were not significantly decreased until 15 hours (Fig. 2A and B; Supplementary Fig. S3A–S3C), and further decreased through 24 hours of drug treatment in the NOTCH1^MUT lines. Furthermore, PI3K inhibition led to a more profound and sustained effect on the levels of these proteins in the NOTCH1^MUT lines than in the NOTCH1^WT lines (Fig. 2B; Supplementary Fig. S3B and S3C). In contrast, although AKT protein levels initially dropped modestly in NOTCH1^WT cells at 4 to 8 hours, AKT levels stabilized at 15 and 24 hours after PI3K inhibition in these cells (Fig. 2A and B; Supplementary Fig. S3A–S3C). Altogether, these data show that apoptosis in NOTCH1^MUT HNSCC cell lines correlates with decreased levels of Aurora A/B, FOXM1, AKT, and/or PDK1 as a result of PI3K pathway inhibition.

We next determined if these proteomic alterations were at the transcriptional level using RNA sequencing (RNA-seq) in two NOTCH1^MUT and one NOTCH1^WT cell lines. We found a significant reduction in AURKA, AURKB, and FOXM1 mRNA levels in all HNSCC cells treated with copanlisib for 18 hours (Fig. 2C). In contrast, PDPK1 (PDK1), AKT1, and AKT2 mRNA levels were unaffected by PI3K inhibition, suggesting that the changes in their protein levels are posttranslational (Fig. 2C).

To gain further insight into how these concordantly regulated proteins interact with the PI3K pathway, we inhibited the individual molecules using kinase inhibitors: SNS-510 (PDK1), MK-2206 (AKT), rapamycin (mTOR), alisertib (Aurora A), and barasertib (Aurora B) at target-specific concentrations. As expected, protein levels of PDK1, AKT, FOXM1, and Aurora A/B decreased substantially following treatment with the dual PI3K/mTOR inhibitor (omipalisib) and pan-PI3K inhibitor (copanlisib) only in NOTCH1^MUT cells (Fig. 2D; Supplementary Fig. S3D–S3E). However, other kinase inhibitors targeting the PI3K pathway, including those affecting PDK1, mTOR, AKT, or PI3Kα (Supplementary Fig. S3E) did not affect levels of Aurora A/B, FOXM1, AKT, or PDK1. Likewise, Aurora A/B kinase inhibition also did not affect levels of any of these proteins (Fig. 2D; Supplementary Fig. S3D and S3E).

Given its differential regulation, we hypothesized that the maintenance of Aurora A/B expression in NOTCH1^WT HNSCC contributed to resistance to PI3K inhibition. To test this hypothesis, we combined the pan-Aurora inhibitor danusertib (0–2 μmol/L) with the dual PI3K/mTOR inhibitor omipalisib (0–200 nmol/L) at a fixed 1:1 ratio in 56 HNSCC cell lines for 72 hours and tested cell viability. The CI was less than 1, indicating synergy, in 46 of 56 HNSCC cell lines (82%) at an effect size of 0.5 and in 49 of 56 cell lines (88%) at an effect size of 0.75 (Fig. 3A). Among the 13 NOTCH1^MUT HNSCC cell lines, all had CI values less than 1 at an effect size of 0.75, suggesting that inhibiting the residual Aurora kinases in NOTCH1^MUT can also enhance cell death.

We also tested the effects of more specific inhibitors of PI3K (copanlisib), Aurora A (alisertib), and Aurora B (barasertib) at clinically relevant concentrations. Alisertib (MLN8237) inhibits catalytic activity of Aurora A, and at higher concentrations can also inhibit Aurora B both in vitro and in vivo (27). Barasertib (AZD2811, AZD1152) is a potent and selective inhibitor of Aurora B (28, 29) that is currently in clinical development (NCT02579226). We treated NOTCH1^WT and NOTCH1^MUT HNSCC cell lines with alisertib, barasertib, or danusertib alone or combined with copanlisib for 24 hours and detected significantly increased induction of apoptosis as measured by cleaved PARP, cleaved caspase-3, and Annexin V and PI staining in the combinations compared with the single agents (Fig. 3B–D Supplementary Fig. S4A–S4D). We found varying sensitivities to Aurora A/B inhibitors alone across all cell lines, with HN31 exhibiting the highest sensitivity and FaDu the lowest. However, the combined inhibition of PI3K and Aurora A/B not only led to increased apoptosis of NOTCH1^MUT HNSCC cell lines but also sensitized otherwise resistant NOTCH1^WT HNSCC cell lines. Furthermore, when we combined barasertib with omipalisib or copanlisib in four NOTCH1^WT cell lines for 72 hours and measured cell viability, the CI values were less than 1 in all four cell lines, indicating synergy (Supplementary Fig. S4E). We used the HEK293 cell line as a non-transformed control and found no significant apoptosis with either single agents or combined PI3K and Aurora kinase inhibition (Supplementary Fig. S4F and S4G). These in vitro findings strongly suggest that combined inhibition of PI3K and Aurora A/B enhanced PI3K-induced apoptosis of NOTCH1^MUT HNSCC cell lines and sensitized NOTCH1^WT HNSCC cell lines.

We then tested these combinations in vivo using a xenograft model of NOTCH1^MUT HNSCC (HN31) and administered copanlisib and alisertib or barasertib for 21 days. When compared with vehicle-treated group (1,466% ± 422%; Fig. 3E; Supplementary Fig. S4H), mice receiving copanlisib or alisertib alone demonstrated significantly smaller tumor volumes (copanlisib: 156% ± 61%, P < 0.05; alisertib: 540% ± 129%, P < 0.05) compared with baseline, whereas the combined treatment led to tumor regression (−78% ± 6%, P < 0.01) at day 19. Similarly, mice treated with a combination of copanlisib and barasertib (−51% ± 18%, P < 0.001; Fig 3F; Supplementary Fig. S4I) showed significant reduction in tumor size, whereas mice receiving copanlisib or barasertib alone exhibited substantial smaller tumor volumes (copanlisib: 187% ± 69%, P < 0.001; barasertib: 223% ± 61%, P < 0.001) when compared with the vehicle treated group (2,182% ± 304%) at day 21. However, alisertib was better tolerated than barasertib and therefore the mice treated with the combination of alisertib and copanlisib underwent a second cycle of treatment and exhibited prolonged and durable tumor regression (Fig. 3E; Supplementary Fig. S4H).

To investigate mechanisms of AR, we exposed the HN31 cell line to increasing concentrations of copanlisib over time until resistance emerged. Subsequently, single-cell clones were established after cell sorting, and the resulting clones [copanlisib acquired resistant (CAR2), CAR10] were tested for sensitivity to copanlisib and omipalisib by cell viability assay and FITC-dUTP/PI staining (Fig. 4A and B; Supplementary Fig. S5A and S5B). The AR clones had a significant shift in IC₅₀ compared with the parental cells (Fig. 4A; Supplementary Fig. S5A) and did not exhibit any significant apoptosis or changes in the cell cycle upon PI3K inhibition (Fig. 4B, C and E; Supplementary Fig. S5B). Moreover, following PI3K/mTOR inhibition, total levels of Aurora B, PDK1, AKT, and FOXM1 decreased more substantially in the parental cells than in the AR clones (Fig. 4C and D; Supplementary Fig. S5C). However, the AR clones still retained copanlisib-induced changes in these proteins, suggesting that they may also engage in additional, novel mechanisms of resistance. To determine whether these protein changes were as a result of changes at the mRNA level, we conducted RNA-seq in the AR clones and HN31 parental cells. Similar to the findings from other NOTCH1^MUT cell lines UMSCC22A and PCI-15B in Fig. 2C, we found > 2 fold reduction in AURKA, AURKB, and FOXM1 mRNA levels in the parental cells but not in the AR clones treated with copanlisib for 18 hours (Supplementary Fig. S5D). Furthermore, the mRNA levels of PDPK1 (PDK1), AKT1, and AKT2 mRNA levels were unaffected by PI3K inhibition in both parental and the AR clones, suggesting that the observed changes in their protein levels are posttranslational (Supplementary Fig. S5D).

To test the hypothesis that mechanisms of innate resistance in NOTCH1^WT HNSCC and AR in NOTCH1^MUT HNSCC may both depend upon Aurora kinases, we inhibited PI3K and Aurora kinases simultaneously in the AR clones and analyzed for apoptosis. We observed increased cleavage of PARP and caspase-3 and higher Annexin V staining (Fig. 4F and G; Supplementary Fig. S5E and S5F) in the cells treated with combined inhibitors compared with vehicle or single-agent copanlisib. Sensitivity to single-agent Aurora kinase inhibitors appeared to be reduced in the AR clones compared with the parental cells, indicating that resistance to PI3K inhibition may alter sensitivity to Aurora kinase inhibitors (i.e., comparing parental HN31 from Fig. 3D with CAR data from Fig. 4G and Supplementary Fig. S5F).

AKT positively regulates the oncogenic transcription factor FOXM1 and phosphorylates a FOXM1 inhibitor (FOXO3a) on a negative regulatory site that was affected after PI3K inhibition only in NOTCH1^MUT cells (Supplementary Fig. S6A). These data led us to hypothesize that the canonical regulation of FOXO3a and FOXM1 is uncoupled from AKT activation in NOTCH1^WT cells, explaining their drug resistance. In support of this hypothesis, FOXM1 can positively regulate genes required for mitosis (30), glycolysis (31, 32), and reactive oxygen species homeostasis (33) so disruption to any or all these processes may contribute to loss of viability. Consistent with this model, significant differences between drug-sensitive NOTCH1^MUT and resistant NOTCH1^WT cell lines were observed in drug-induced levels of key enzymes regulating glucose metabolism and cellular redox homeostasis (Supplementary Fig. S1B). To test this hypothesis, we manipulated levels of FOXM1 and scavenged reactive oxygen species in HNSCC cells treated with omipalisib and determined the effects on cell survival. Knockdown of FOXM1 in NOTCH1^WT cells did not sensitize them to omipalisib-mediated apoptosis (Supplementary Fig. S6B). Overexpression of FOXM1 did not rescue NOTCH1^MUT cells from omipalisib-induced apoptosis (Supplementary Fig. S6C and S6D). Although reactive oxygen species increased after treatment with omipalisib in NOTCH1^MUT cells and was effectively scavenged by N-acetyl cysteine, treatment with N-acetyl cysteine did not rescue apoptosis in NOTCH1^MUT cells (Supplementary Fig. S6E–S6G).

It was previously reported that Aurora A contributes to resistance to PI3K inhibition in breast cancer (34). We examined this possibility in our HNSCC models by overexpressing Aurora A in NOTCH1^MUT cells. Subsequent treatment with copanlisib for 24 hours did not reverse the apoptotic phenotype of NOTCH1^MUT cells, suggesting that an alternate pathway is responsible for drug sensitivity (Supplementary Fig. S6H).

Because Aurora A and FOXM1 did not modulate sensitivity to PI3K inhibition in NOTCH1^MUT HNSCC, we then tested the effect of altering Aurora B. Overexpression of Aurora B in NOTCH1^MUT HNSCC cells partially rescued copanlisib-induced apoptosis, as demonstrated by markedly reduced cleaved PARP and caspase-3 proteins and Annexin V positive cells (Fig. 5A and B). AKT and PDK1 levels were also upregulated when Aurora B was overexpressed in NOTCH1^MUT HNSCC (Fig. 5A). In NOTCH1^WT HNSCC cells, siRNA-mediated Aurora B knockdown significantly enhanced PI3K inhibition–induced apoptosis (Fig. 5C and D; Supplementary Fig. S7A–S7C), but not to the same extent as we observed in NOTCH1^MUT HNSCC (HN31) cells despite a similar level of Aurora B expression following treatment with copanlisib (Fig. 5C). Likewise, total levels of AKT and PDK1 were not significantly affected by Aurora B knockdown in NOTCH1^WT HNSCC cells (Fig. 5C; Supplementary Fig. S7A).

Because total AKT levels were upregulated upon Aurora B overexpression in NOTCH1^MUT HNSCC cells with concordant protection from PI3K inhibition–induced apoptosis, we examined the effect of total AKT on PI3K inhibition–induced apoptosis. We manipulated AKT1, the predominant isoform of AKT. Overexpression of AKT1 in NOTCH1^MUT HNSCC cells significantly reduced PI3K inhibition–induced apoptosis (Fig. 5E and F). Moreover, we observed corresponding changes in the PDK1 protein levels but not Aurora B levels upon AKT1 overexpression in NOTCH1^MUT HNSCC cells (Fig. 5E). In addition, AKT1 knockdown in NOTCH1^WT HNSCC cells treated with PI3K inhibition resulted in markedly higher cell death (Fig. 5G and H; Supplementary Fig. S7C and S7D).

These findings illustrate that total Aurora B governs the expression of total AKT, which subsequently regulates PDK1 levels in NOTCH1^MUT HNSCC cells, in which Aurora B, AKT, and PDK1 are crucial effectors that determine cell survival in response to PI3K inhibition. However, in NOTCH1^WT HNSCC cells, Aurora B does not regulate total AKT or PDK1 levels (Fig. 5I).

To address the need for biomarker-selected targeted therapy for HNSCC, we previously demonstrated that PI3K inhibition caused apoptosis selectively in NOTCH1^MUT HNSCC, but the mechanisms of resistance in NOTCH1^WT HNSCC were unknown. In this study, we show that PI3K inhibition leads to reduced Aurora B levels, which in turn regulate total AKT protein levels exclusively in NOTCH1^MUT HNSCC in a kinase-independent manner. Subsequently, AKT affects PDK1 protein levels, also independent of kinase activity. Total AKT and PDK1 loss mediates apoptosis after PI3K inhibition in NOTCH1^MUT HNSCC. In addition, the pathways involving maintenance of protein levels of Aurora kinases in response to PI3K inhibition contribute to both innate resistance and AR to PI3K inhibition in HNSCC. In our study, concurrent Aurora kinase and PI3K inhibition led to increased cell death in vitro and in vivo. In contrast, reduced Aurora A and FOXM1 levels were associated with, but did not control the apoptotic phenotype in NOTCH1^MUT HNSCC cells.

The mechanism we propose identifies several previously unrecognized interactions between the PI3K/AKT pathway and Aurora kinases. Our discovery raises several questions for future studies that would dissect specific interactions within this pathway, including understanding how Aurora B levels are differentially regulated in NOTCH1^WT and NOTCH1^MUT HNSCC in response to PI3K inhibition. Because Aurora B levels are altered at both mRNA and protein levels, it is important to determine the molecular factors responsible for these changes. One potential mediator may be FOXO3a. We observed significantly lower levels of phosphorylated FOXO3a in response to PI3K inhibition in NOTCH1^MUT HNSCC cells than in NOTCH1^WT cells. FOXO3a undergoes AKT-mediated phosphorylation at S235 and is rendered inactive, thus being unavailable to bind to the promoters of its numerous targets, including Aurora B, and repress their transcription (35). Involvement of FOXO3a may contribute to the regulation of AURKB mRNA levels. For identification of posttranscriptional regulators of Aurora B, it will be imperative to determine the protein half-life of Aurora B in NOTCH1^WT and NOTCH1^MUT HNSCC cells with and without PI3K inhibition.

Another striking finding from the current study is the sustained, depleted protein levels of total AKT in NOTCH1^MUT HNSCC in response to PI3K inhibition. Unlike Aurora B, AKT is clearly regulated at the posttranscriptional level. Moreover, the effect of Aurora B protein expression on AKT is not solely dependent upon kinase activity. Therefore, it will be intriguing to investigate how Aurora B and AKT interact with each other in a kinase-independent manner. Furthermore, our time-course studies showed that AKT protein levels initially decrease in NOTCH1^WT and NOTCH1^MUT cells with more marked, durable changes in NOTCH1^MUT. Further work is warranted to understand the mechanism behind AKT protein downregulation. One potential candidate to mediate this differential effect is BRCA1. BRCA1 mutant cells accumulate nuclear phospho-AKT and consequently inactivate the transcription functions of FOXO3a, a main nuclear target of phospho-AKT (36). In addition, NOTCH1 activation further compensates for BRCA1 deficiency and promotes survival of triple-negative breast cancer (37). Furthermore, BRCA1 phosphorylation is regulated by the PI3K pathway, and thereby its subcellular localization and functions (38). In summary, the differential sensitivity to PI3K inhibition in NOTCH1^WT and NOTCH1^MUT HNSCC could be due to the differential activation of BRCA1, which could also explain the proteomic alterations in AKT and Aurora B (39).

Another aspect that remains to be understood is how total AKT1 regulates PDK1 levels in NOTCH1^MUT but not in NOTCH1^WT HNSCC. A recent study showed that CK1- and GSK3β-mediated phosphorylation of PDK1 led to its ubiquitination and degradation by E3 ubiquitin ligase speckle type BTB/POZ protein (SPOP; ref. 40). We speculate that increased AKT levels lead to inactivation of GSK3β, which then fails to phosphorylate PDK1, leading to its degradation by SPOP. It is possible that knockdown of AKT1 alone in NOTCH1^WT HNSCC does not alter PDK1 to a significant extent by itself, because the other AKT isoforms could function as redundant proteins in this context. However, when the PI3K pathway is inhibited in AKT1 knockdown cells, activated GSK3β could potentially mediate PDK1 degradation in NOTCH1^WT HNSCC. This mechanism could potentially explain our observation of differential PDK1 levels upon AKT1 manipulation in NOTCH1^MUT and NOTCH1^WT HNSCC in response to PI3K inhibition.

We showed that the combination of PI3K and Aurora kinase inhibition is synergistic and leads to increased apoptosis in most HNSCC cells independent of mutation status, including those with innate and AR to PI3K inhibitors. These in vitro findings were validated in our in vivo models, which showed robust tumor regression in mice receiving combined therapy. Our findings suggest that this combination would be broadly effective against HNSCC in patients who may have heterogeneous tumors. In addition, it is rational to target a pathway that mediates AR initially to achieve a more durable response to therapy (41).

The mechanism that underlies the synergy between PI3K and Aurora kinase inhibition is likely distinct from the model we propose to explain the resistance of NOTCH1^WT HNSCC to PI3K inhibitors, which is independent of Aurora A and the kinase activity of Aurora B. One possible explanation for the synergy is that because PI3K inhibition leads to reduced total levels of Aurora kinases, these cells are more dependent on the remaining activity of Aurora kinases for mitotic progression. In addition, prolonged inhibition of Aurora A can lead to inhibition of Aurora B (27). A second possible explanation for the synergy hinges on the finding that PI3K inhibition leads to decreased Rb protein expression in both NOTCH1^MUT and NOTCH1^WT HNSCC cell lines. Two independent studies have shown that cancer cells with loss of RB1 are hyper-dependent on Aurora A and Aurora B for survival (42, 43).

Donnella and colleagues demonstrated a decrease in total AURKA mRNA and protein 24 hours following PI3K inhibition in sensitive breast cancer cell lines (34). The combination of a PI3K inhibitor with alisertib was synergistic in 38% of breast cancer lines. They demonstrated that MYC-driven AURKA expression maintains AKT and mTOR activity; inhibition of Aurora A enhances PI3K inhibition by contributing to the complete suppression of AKT/mTOR signaling. In contrast to our model in HNSCC, they found that AURKB was not significantly associated with sensitivity to PI3K inhibitors in breast cancer.

The mechanisms of AR to PI3K inhibition in NOTCH1^MUT HNSCC could be driven by additional mechanisms that do not overlap with mechanisms of innate resistance in NOTCH1^WT HNSCC. As is the case with targeted therapies in non–small cell lung cancer, there are several distinct mechanisms of AR in EGFR mutant and anaplastic lymphoma kinase—positive subsets. They either involve on-target mechanisms involving gene/target amplification that enables continuous downstream signaling or off-target effects, which result in activation of bypass signaling (44).

Because both PI3K (e.g., paxalisib, umbralisib, parsaclisib, copanlisib, and duvelisib) and Aurora kinase inhibitors (alisertib, barasertib) are in clinical development, our work could be rapidly translated to clinical testing. Alternatively, AKT and Aurora kinase–specific proteolysis-targeting chimeras, which are in the process of development and validation, might be an effective therapeutic option in cases where PI3K inhibitors fail (25, 45, 46). Notably, our recent clinical trial testing a dual PI3K/mTOR inhibitor in patients with NOTCH1^MUT HNSCC with recurrent or metastatic disease showed modest single-agent clinical activity (NCT03740100), indicating that combination therapy could be an effective approach (8, 47).

These findings collectively show that sustained Aurora B expression via AKT and PDK1 levels drives resistance to PI3K inhibition–induced apoptosis in NOTCH1^WT HNSCC. We have defined a mechanism that drives sensitivity and resistance to PI3K inhibitors in NOTCH1^MUT HNSCC and propose combined PI3K and Aurora kinase inhibition to maximize clinical efficacy and overcome innate and AR to PI3K inhibitors, thereby establishing a foundation for future clinical trials.

F.M. Johnson reports a patent for UTSC.P1483US.CP1 pending to MD Anderson Cancer Center, and F.M. Johnson has received research funding from Takeda and Viracta Therapeutics within the last 36 months. No disclosures were reported by the other authors.

P.A. Shah: Conceptualization, formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. V. Sambandam: Conceptualization, visualization, methodology, writing–original draft. A.M. Fernandez: Investigation, visualization. H. Zhao: Investigation. T. Mazumdar: Investigation, writing–original draft. L. Shen: Formal analysis, visualization. Q. Wang: Formal analysis, visualization. K.M. Ahmed: Investigation. S. Ghosh: Writing-review and editing. M.J. Frederick: Conceptualization, funding acquisition, methodology, project administration, writing–review and editing. J. Wang: Formal analysis. F.M. Johnson: Conceptualization, supervision, funding acquisition, methodology, writing–original draft, project administration, writing–review and editing.

The authors thank Erica Goodoff in MD Anderson's Research Medical Library for editing the manuscript. This work was supported by philanthropic contributions to The University of Texas MD Anderson Cancer Center's Oropharynx Discovery Program (to F.M. Johnson), from the NIH (1R01CA235620, to F.M. Johnson and M.J. Frederick), and the Cancer Prevention and Research Institute of Texas (RP200369 to F.M. Johnson and M.J. Frederick). Flow cytometry and bioinformatics analyses were supported by the NCI through MD Anderson's Cancer Center Support Grant (P30CA016672).

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.