Integrative Kinome Profiling Identifies mTORC1/2 Inhibition as Treatment Strategy in Ovarian Clear Cell Carcinoma

This article is featured in Highlights of This Issue, p. 3785

In the United States, ovarian cancer is the fifth-leading cause of cancer deaths in women (1). Ovarian clear cell carcinoma (OCCC) is the second most common subtype of epithelial ovarian cancer. The majority of patients with OCCC are diagnosed at an early stage (57%–81% at stage I/II) and have better overall survival compared with stage-matched high-grade serous (HGS) ovarian cancer, the most common subtype of ovarian cancer. In contrast, patients with OCCC diagnosed at late stage respond poorly to standard platinum-based chemotherapy compared with patients with late-stage HGS ovarian carcinoma (2). In recent years, genetic studies in relatively small patient groups have revealed the mutational landscape in OCCC. The SWI-SNF chromatin remodeling complex DNA-binding AT-rich interactive domain 1A gene (ARID1A) has been shown to be deleteriously mutated in 40% to 57% of patients with OCCC, the highest percentage found in any cancer (3, 4). Loss of ARID1A protein, being a key component of the complex, may affect the expression of many genes (5). Activation of the PI3K/AKT/mTOR pathway, implicated in survival, protein synthesis, and proliferation, is another major player in OCCC. PIK3CA, encoding the catalytic domain of PI3K, contains activating mutations in 30% to 40% of patients with OCCC, whereas expression of the PI3K antagonist PTEN is diminished in 40% of patients with OCCC (4, 6, 7). Furthermore, mutations in the oncogene KRAS and the tumor-suppressor gene TP53 have been identified in 4.7% to 14% and 10% to 15% of patients with OCCC, respectively (1, 4, 6, 8, 9). In addition to mutational aberrations, copy-number alterations (CNAs) have been found in OCCC tumor samples in the proto oncogene ZNF217, tumor-suppressor genes, cyclin-dependent kinase inhibitors CDKN2A and CDKN2B, and the membrane receptor oncogene MET (10–12).

The identification of the most frequently mutated genes ARID1A and PIK3CA may lead to new therapeutic strategies. In particular, the effects of ARID1A loss are being investigated, and vulnerabilities in ARID1A-mutant cancers are being identified. Synthetic lethal interactions have recently been demonstrated in ARID1A-mutant OCCC cancer cell lines by shRNA-mediated suppression of ARID1B, a homolog of ARID1A, as well as chemical inhibition of the histone H3 methyltransferase EZH2 and histone deacetylase HDAC6 (13–15). PI3K signaling–mediated tumor addiction through the well-studied hypermorphic mutant forms of PIK3CA (E545* and H1047*) was studied extensively in multiple cancer types including OCCC. Recent translational research in OCCC cell lines demonstrated sensitivity to PI3K/mTOR dual inhibitors and AKT inhibitors, although PIK3CA mutations did not predict sensitivity to these inhibitors (16, 17).

In the present study, we aimed to identify novel targetable mutations by means of high-coverage sequencing of all protein kinase genes, referred to as the kinome, and of a subgroup of cancer-related genes in a large set of OCCC. In addition, we determined copy-number gains and losses in kinases and other genes of OCCC tumors using high-coverage single-nucleotide polymorphism (SNP) arrays. To detect kinase mutations and CNA at both high and low frequencies, we used a large cohort of 124 untreated primary OCCC tumors and most of the available OCCC cell lines (n = 17). Finally, we functionally validated several candidate targets in OCCC cell lines and unique OCCC patient-derived xenograft (PDX) models. Our results indicate mTORC1/2 inhibition as an approach to guide future development of therapeutic strategies for OCCC.

Primary tumor samples from 124 patients with OCCC and 47 paired control blood samples were prospectively collected from Belgium, Germany, Norway, Poland, the Netherlands, the UK, and the United States. All patients gave written informed consent for samples to be collected, and the corresponding ethical review boards approved the study. Tumor samples had to contain ≥40% tumor cells, of which ≥70% was OCCC, as determined by experienced gynecologic oncology pathologists. We obtained 17 human OCCC cell lines: TOV21G (ATCC); RMG1, RMG2, OVMANA, OVTOKO, and HAC2 (JCRB Cell Bank); JHOC5 (RIKEN Cell Bank); OVCA429 (Cell Biolabs); OVSAYO, TUOC1, KK, OVAS, SMOV2, and KOC7C (Dr. Hiroaki Itamochi, Tottori University School of Medicine, Tottori, Japan); ES2 (Dr. Els Berns, Erasmus MC, Rotterdam, the Netherlands); TAYA (Dr. Yasushi Saga, Jichi Medical University, Yakushiji, Shimotsuke-shi, Tochigi, Japan); and OV207 (Dr. Vijayalakshmi Shridhar, Mayo Clinic). All cell lines were maintained in RPMI supplemented with 10% FCS, 100 μg/mL Penicillin/Streptomycin and 2 mM L-glutamine. All cell lines were tested by short tandem repeat profiling and were mycoplasma free. All cell lines were kept in culture for a maximum of 50 passages.

From 124 primary fresh-frozen OCCC tumors and 47 paired controls, 3-μg DNA was prepared for sequencing using the following steps. Genomic DNA was sheared to produce 300 bp fragments (Covaris S220); using SureSelect Target enrichment & Human Kinome Kit (Agilent technologies), kinase exons were tagged and captured; using biotinylated RNA library baits and streptavidin beads, exons were amplified and loaded on a HiSeq2500 Illumina sequencer using paired-end sequencing according to the manufacturer's protocols. The SureSelect Human Kinome Kit captures exons from 518 kinases, 13 diglyceride kinases, 18 PI3K domain and regulatory component genes, and 48 cancer-related genes (Supplementary Table S1). After sequencing, raw data were mapped to the human reference sequence NCBI build 37 (hg19) and processed according to our sequencing pipeline (Supplementary Fig. S1). Genome Analysis Toolkit (GATK, version 1.0.5069) was used for indel realignment and base quality recalibration on BAM files. See the Supplementary Methods for further details on kinome sequencing.

Genome-wide SNP genotyping was performed with HumanOmniExpressExome-8BeadChip (Illumina) containing >900K SNPs, including >273K functional exomic markers to determine CNA in 108 primary OCCC tumors and 17 OCCC cell lines. DNA sample processing, hybridization, labeling, scanning, and data extraction were performed according to Illumina infinium 2 protocol. Illumina GenomeStudio software was used for primary sample assessment and SNP call rate quality control of SNP intensity output files. See the Supplementary Methods for further details on SNP array analysis.

The 17 OCCC cell lines were plated in 384-well plates with 250 to 2,000 cells/well using an automatic cell dispenser. Cell lines were incubated with varying concentrations of AZD8055 (Axon Medchem), gefitinib (Selleckchem), lapatinib (Selleckchem), or olaparib (Axon Medchem) for 6 days. Cell viability was subsequently monitored using CellTiter-Blue cell viability assay (Promega). Due to large-cell density-dependent variability in drug sensitivity, for each cell line, the lowest and highest IC₅₀ were removed. The cell lines CAPAN1, OE19, LOVO, OVCAR3, OVCAR4, and SKOV3 were exposed to AZD8055 to serve as sensitive and resistant controls.

All animal experiments were approved by the Institutional Animal Care and Use Committee of the University of Groningen (Groningen, the Netherlands) and carried out in accordance with the approved guideline “code of practice: animal experiments in cancer research” (Netherlands Inspectorate for Health Protection, Commodities and Veterinary Public Health, 1999). See the Supplementary Methods for further details on AZD8055 in vivo administration.

K-means consensus clustering of kinome mutations and CNA, mRNA level determination, flow cytometry, MTT assays, long-term proliferation assays, Western blotting, immunohistochemical analysis, data visualization, and statistical analysis are described in the Supplementary Methods.

Across all OCCC tumors (n = 122) and OCCC cell lines (n = 17), 95.9% of all bases had >20 read coverage for variant calling. The mean coverage depth for aligned reads was 99.3x. On average, 1.17% of the sequenced genes were mutated per patient. Genes with a high mutation frequency (>4%) across all OCCC tumors are shown in Fig. 1A. Resequencing of these 19 genes using Haloplex confirmed 227 of the 234 mutations (97%) that were originally identified by kinome sequencing using the same tumor DNA. The majority of genes with a high mutation frequency are implicated in well-known cancer-related pathways like the PI3K/AKT/mTOR pathway (PIK3CA, PTEN, PIK3R1, and AKT1), MAPK signaling transduction pathway (KRAS), DNA repair pathway (TP53, ATM, and PRKDC), ERBB family of receptor tyrosine kinase genes (ERBB3 and EGFR), and chromatin remodeling genes (ARID1A). The frequencies of previously described mutated genes observed in our study were in agreement with such frequencies reported earlier in smaller studies of OCCC (Fig. 1A; refs. 3, 4, 6, 8, 9).

ARID1A mutations were analyzed using Haloplex sequencing only. Of 54 ARID1A-mutant tumors, 18 tumors contained homozygous frameshift or stop-gain mutations, 13 tumors harbored more than one heterozygous mutation, whereas 23 tumors contained a single heterozygous frameshift or stop-gain mutation (Fig. 1B). The identified proportion of each type of mutation in ARID1A matches those reported in earlier studies (3, 4). ARID1A-mutant and wild-type tumors did not show differential mutation incidence in kinome genes (7.4 vs. 6.6 mutations per tumor on average).

Statistical binomial univariate testing of all 851 identified kinome mutations revealed 11 significantly mutated genes (P < 0.05) relative to background mutations. These predicted oncogenic drivers were PIK3CA, KRAS, TP53, PTEN, AKT1, PIK3R1, FBXW7, ERBB3, ATM, CHEK2, and MYO3A (Supplementary Table S3). PIK3CA and KRAS exhibited mutations in established hotspot sites in aa Q542, Q545, and H1047 in PIK3CA and G12 in KRAS, whereas TP53 mutations were distributed across the TP53 DNA-binding domain (Fig. 1C–E).

Interestingly, we identified mutations in three genes not previously described in OCCC (AKT1, PIK3R1, and ERBB3) and with an established role in PI3K/AKT/mTOR pathway activation. AKT1 missense mutations were identified in six tumors (4.9%). AKT1 is one of the key components of the PI3K/AKT/mTOR cascade. The PH domain of AKT1 interacts with its kinase domain and maintains the protein in a closed and inactive state (18). One candidate somatic mutation, D323N, was located in the kinase domain. The mutations R25H, L52R (occurring in three tumors), and W80R (both described in COSMIC) were all located in the AKT1 PH domain (Supplementary Fig. S2A). PIK3R1 was mutated in nine tumors (7.3%), of which seven tumors carried mutations in the inter-SH2-1–SH2-2 domain of the protein (aa 429–623). This domain binds PIK3CA and is required for the inhibitory role of PIK3R1 on PIK3CA (19). Two of the inter-SH2 domain mutations are described in COSMIC: E439* and T576* in-frame deletions. In addition, we identified six novel inter-SH2 domain mutations; two somatic missense (aa N453D and T471S, together in one tumor) and two somatic in-frame deletions (aa QF455* and RE461*), one candidate somatic frameshift (starting in aa 582), and one candidate somatic stop-gain mutation (aa 571; Supplementary Fig. S2B). In ERBB3, we identified 10 missense mutations across eight distinct tumors (5.7%). At the cell membrane, ERBB3 can dimerize with other ERBB family members and regulate downstream kinase signaling. Six mutations were located across the extracellular domains of ERBB3, and two could be identified as somatic. The D297Y mutation has been described as a hotspot location in ovarian and colorectal cancers (20). Furthermore, three mutations were in the intracellular C-terminal domain and one in the kinase domain (Supplementary Fig. S2C). Of the other significantly mutated genes, F-Box and WD Repeat Domain Containing 7 (FBXW7) has been assigned a role in PI3K regulation and DNA repair. These mutations are mainly found in its WD repeats (21). ATM Serine/Threonine Kinase (ATM) and Checkpoint kinase 2 (CHEK2) are designated as DNA repair genes, whereas Myosin IIIA (MYO3A) is an actin-dependent motor protein. Mutations in FBXW7 and ATM were previously described in OCCC (Supplementary Fig. S2D–S2G; ref. 22).

We also investigated the pathways in which genes with a high mutation frequency are involved. To this end, all low-mutation frequency genes (<4%) present in PI3K/AKT/mTOR, MAPK, and DNA repair pathways, the ERBB family of receptor tyrosine kinases and mutations in ARID1A were mapped per tumor in order to determine mutational spectra across these signaling pathways in OCCC. One or more mutations were found in PI3K/AKT/mTOR pathway genes in 76 patients (62.3%), in MAPK pathway genes in 22 patients (18%), in the ERBB family of receptor tyrosine kinases in 18 patients (14.8%), and in DNA repair pathway genes in 46 patients (37.7%). The co-occurrence of mutations in both ARID1A and PIK3CA in this large set of OCCC tumors was in accordance with the findings from a previous, smaller study (Supplementary Fig. S3A; refs. 4, 23). PIK3R1 mutations were mutually exclusive with PIK3CA-mutant tumors (P = 0.043), and a trend was observed for PTEN with PIK3CA-mutant tumors (P = 0.0756). In general, TP53-mutant tumors (including pure OCCCs) were mutually exclusive with both ARID1A- and PIK3CA-mutant tumors (P = 0.0031), in agreement with previous literature (5). Surprisingly, co-occurrence of TP53 and ARID1A or PIK3CA mutations was observed in a few cases. In total 108 of 122 tumors (89%) comprised one or more mutations in genes belonging to the PI3K/AKT/mTOR pathway, MAPK pathway, DNA repair pathway, and ERBB family of receptor tyrosine kinases (Supplementary Fig. S3A).

Evidently, there is a large overlap in tumors that were both ARID1A mutant and harbored a mutation in DNA repair genes or PI3K/AKT/mTOR, MAPK, and ERBB family of receptor tyrosine kinase genes (Supplementary Fig. S3B). Subdividing tumors into ARID1A heterozygous-mutated and homozygous-mutated tumors did not change the overlap in mutations in PI3K/AKT/mTOR, MAPK, and ERBB family of receptor tyrosine kinase genes or DNA repair genes in both groups (data not shown). ARID1A wild-type tumors more frequently contained mutations in DNA repair genes in addition to mutations in PI3K/AKT/mTOR, MAPK, or the ERBB family of receptor tyrosine kinase genes as compared with ARID1A-mutant tumors (49% vs. 26%, P = 0.0385).

In 108 SNP-genotyped primary OCCC tumors, GISTIC (Genomic Identification of Significant Targets in Cancer) analysis annotated 324 significantly amplified genes located in 48 focal regions. We identified amplification of ZNF217 (20q13.20), a transcriptional regulator previously described in OCCC (11), in 29 tumors (27%, P = 0.014). Using GISTIC analysis, we subsequently found multiple kinases and other cancer-related genes among the significantly amplified genes (Fig. 2).

One recurrently amplified region (3q26.2) contained multiple kinases: PIK3CA, EPHB3, MAPK3K13, PAK2, PRKCI, TNIK, TNK2, and DGKG (14 tumors, 13%, P = 0.046). The cancer-related gene TERT (5p15.33), included in kinome sequencing, was amplified in 12 tumors (11%, P = 0.044). EGFR (7p11.2) emerged as the most significant recurrently amplified kinase; it was present in 18 OCCC tumors (17%, P < 0.0001). Other recurrently amplified kinases included ERBB2 (17q12), which was amplified in 21 tumors (19%, P = 0.002), and the chromatin-associated and transcriptional control-related kinase TRIM28 (19q13.43), which was amplified in 52 tumors (48%, P = 0.0035).

Furthermore, a total of 118 significantly deleted genes were identified in 62 focally deleted regions. Two significantly deleted kinases could be identified: SIK3 (11q23.3) and the transcriptional repressor CDK8 (13q12.13; Fig. 2). In addition, 3 known cancer-related genes included in kinome sequencing were significantly deleted. The cell-cycle regulators CDKN2A and CDKN2B (both located on 9p21.3) were deleted in 16 tumors (15%, P = 0.024), and PALB2 (16p12.2), an essential chaperone of BRCA2, was deleted in 33 tumors (31%, P < 0.0001).

A separate GISTIC analysis was implemented to compare ARID1A-mutant tumors (n = 45) with ARID1A wild-type (n = 63) tumors (Supplementary Fig. S4A). TRIM28 amplification and CDK8 deletion were significantly retained only in ARID1A-mutant tumors (P < 0.0001 and P = 0.0067, respectively), whereas EGFR amplification was significantly retained only in ARID1A wild-type tumors (P = 0.0054; Supplementary Fig. S4B and S4C).

Both kinome sequencing and SNP data were available for 106 tumors. After merging all mutations and CNA events in the PI3K/AKT/mTOR, MAPK, and DNA repair pathway and ERBB family of receptor tyrosine kinases, we identified at least one event in 103 of 106 tumors analyzed (97%; Fig. 3A).

Amplification incidence of >10% was found in the PI3K/AKT/mTOR-related genes AKT2 (17%), PIK3R2 (14.2%), PIK3CA (12.3%), AKT1 (10.4%), the ERBB family of receptor tyrosine kinases ERBB2 (19.8%) and EGFR (17%), and the DNA repair genes PRKDC (24%) and CHEK2 (12.3%). AKT2, PIK3R2, ERBB2, EGFR, and CHEK2 primarily contained amplifications, from which only ERBB2 (n = 2) and EGFR (n = 1) presented tumors that were both amplified and mutated. PIK3CA was mostly mutated, yet 4 PIK3CA-mutant tumors carried both an amplification and a mutation. Amplifications and mutations never co-occurred in AKT1 and PRKDC.

Deletion incidence of >10% was observed in the PI3K/AKT/mTOR-related gene PIK3C3 (18.9%) and the DNA repair genes PALB2 (30.2%), ATM (17.9%), and BRCA2 (17%). PIK3C3, PALB2, and BRCA2 primarily harbored deletions, whereas ATM deletions and mutations co-occurred in three tumors.

We hypothesized that the alterations in genes described in Fig. 3A can be added up to promote aberrant pathway signaling. Mutations or CNA occurred in PI3K/AKT/mTOR in 84.9% of tumors, in MAPK pathway in 27.4% of tumors, and in ERBB receptor family of kinases in 42.5% of tumors. Combined mutations and CNA in PI3K/AKT/mTOR and MAPK pathway genes and the ERBB family of receptor tyrosine kinases indicated that 91% of all tumors were affected (Fig. 3B), whereas the DNA repair pathway was affected in 82% of all tumors (Fig. 3C).

Clinical data were available for a subset of patients (n = 70; Supplementary Table S4). Disease-specific survival analysis revealed that ARID1A, PIK3CA, or ARID1A plus PIK3CA alterations were not related to survival (Supplementary Fig. S5A–S5C). Kinome mutations and CNA events of all 106 tumors were integrated, and the tumors were grouped using K-means consensus clustering. Maximum cluster number was set at 8, because more clusters only marginally decreased friction (Supplementary Fig. S5D–S5F). Most tumors grouped together in cluster 1 (n = 53), 3 (n = 27), or 5 (n = 13). A trend for worse disease-specific survival was demonstrated with cluster 3 compared with all other clusters (P = 0.0638; Supplementary Fig. S5G and S5H), which became highly significant upon selection for patients with advanced-stage OCCC (P < 0.001, n = 31; Supplementary Fig. S5I). Grouping of mutation and CNA status for each cluster did not reveal unique genes in cluster 3 (Supplementary Fig. S5J and S5K).

Kinome sequencing and SNP data of 17 OCCC cell lines presented similar alteration frequencies in PI3K/AKT/mTOR, MAPK, and DNA repair pathway genes and the ERBB family of receptor tyrosine kinases as those observed in patients with OCCC (Fig. 4A and B). However, an overrepresentation of ARID1A and TP53 mutations and an underrepresentation of KRAS mutations were found. Due to the resemblance of OCCC patient alterations with those in the cell line panel, we decided to screen for kinase inhibition vulnerabilities of the PI3K/AKT/mTOR and MAPK pathway downstream targets mTORC1/2 and the ERBB receptor tyrosine kinases EGFR and ERBB2. In addition, the DNA repair pathway was targeted. High-throughput drug sensitivity testing revealed nanomolar range efficacy against the mTORC1/2 inhibitor AZD8055 in all 17 OCCC cell lines tested (Fig. 5A). Sensitivity appeared to be irrespective of PIK3CA or ARID1A mutation status, suggesting that alterations upstream of mTORC1/2 may explain the comprehensive AZD8055 susceptibility. COSMICs cancer cell line drug screening data (n = 913 cell lines) with AZD8055 (Cancerrxgene.org) and temsirolimus (targeting mTORC1) indicate that OCCC cell lines are among the most sensitive ovarian cancer cell lines (Fig. 5B and Supplementary Fig. S6A).

MTT assay–based IC₅₀ determination in 14 of 17 OCCC cell lines demonstrated higher susceptibility to AZD8055 and the mTORC1 inhibitor everolimus in comparison with 3 HGS ovarian cancer cell lines (Fig. 5C). In contrast to AZD8055, proliferation inhibition by everolimus displayed a plateau phase in a wide concentration range and a large variation in sensitivity among OCCC cell lines (Fig. 5D and Supplementary Fig. S6B). Moreover, everolimus treatment resulted in increased p-AKT⁴⁷³ levels, whereas AZD8055 and MLN0128, another mTORC1/2 inhibitor, reduced this upregulation (Supplementary Fig. S6C). AZD8055 and dactolisib, an mTORC1/2-PI3K inhibitor for which OCCC cell lines are highly sensitive (Supplementary Fig. S6A), strongly reduced p-AKT⁴⁷³ and to a lesser extent p-AKT³⁰⁸ at high concentrations (Supplementary Fig. S6D). The absence of PARP cleavage indicated that everolimus, AZD8055, and dactolisib do not induce apoptosis in an in vitro setting (Supplementary Fig. S6D). OCCC cell lines presented a large IC₅₀ variation for upstream inhibitors of the PI3K/AKT/mTOR and MAPK pathway, for example, PI3K inhibitor GDC0941 and MEK inhibitor selumetinib, compared with AZD8055 and dactolisib (Supplementary Fig. S6E). Long-term proliferation inhibition by dactolisib was stronger compared with AZD8055 (Supplementary Fig. S6F). Limited sensitivity was observed toward the EGFR inhibitor gefitinib and ERBB2/EGFR inhibitor lapatinib. EGFR mutations were observed in two cell lines, but only OVMANA (R705K) displayed gefitinib sensitivity. In cells with elevated levels of ERBB2 mRNA and high ERBB2 membrane expression (RMG1, SMOV2, and TUOC1), lapatinib sensitivity was observed (Fig. 5A). However, some of the tested cell lines without EGFR mutations or elevated ERBB2 levels also displayed sensitivity to gefitinib and lapatinib, indicating involvement of alternative mechanisms. The PARP1 inhibitor olaparib showed some efficacy against ATR-mutant TOV21G cells.

In conclusion, OCCC cell lines exhibit exquisite sensitivity to mTORC1/2 inhibitors, whereas limited sensitivity was observed toward ERBB receptor tyrosine kinase and DNA repair inhibitors despite the high aberration frequency in those pathways.

Considering the high susceptibility of OCCC cell lines to AZD8055, we validated mTORC1/2 activity in OCCC tumor samples. Immunostaining of the mTORC1/2 downstream target p-S6 was performed on two different tissue microarrays (n = 136 and n = 83) with primary tumor material from HGS ovarian cancer and OCCC. OCCC tumors were more frequently p-S6 positive compared with HGS ovarian cancer in both datasets (P = 0.053 and P = 0.0028, respectively; Fig. 6A–C). Subsequently, we tested AZD8055 efficacy in OCCC PDX–bearing NSG mice. Sequencing of 19 genes with the highest mutation frequency in patients with OCCC showed mutations in ARID1A (1148* stop-gain), PTEN (H93R), and BRCA1 (V356A and 1533* stop-gain) in PDX.155; in PIK3CA (K111E, recurrent in OCCC tumors) and ATM (T1020I) in PDX.180; and none in PDX.247. All mutations were present in the sequenced F0 (patient tumor), F1, and F2 generation. GISTIC analysis indicated CNA across the PDX models that resemble CNA in patients with OCCC (Supplementary Fig. S7A). Significant growth differences were observed between AZD8055 (10 mg/kg/day) and vehicle-treated mice in all 3 PDX models (Fig. 6D–F). AZD8055 treatment reduced tumor growth in PDX.247 and inhibited tumor growth in PDX.155 and PDX.180, which is reflected in a reduction of proliferation marker Ki67 positivity (Fig. 6D–F). Remarkably, an increased staining of the apoptosis marker Cleaved Caspase-3 was observed in PDX.155 after AZD8055 treatment (Fig. 6D–F). Phosphorylation of mTORC1/2 downstream target S6 was clearly detectable in the vehicle-treated tumors but did not significantly change after AZD8055 treatment (Supplementary Fig. S7B–S7D).

Patients with advanced-stage OCCC respond poorly to standard platinum-based chemotherapy, indicating an unmet need for novel treatment strategies. Kinome sequencing of 518 kinases and 79 cancer-related genes and SNP array analysis of 108 OCCC tumors revealed mutations, amplifications, and deletions in well-known genes such as ARID1A, PIK3CA, TP53, and KRAS and several newly identified genes. Collectively, alterations in genes from the PI3K/AKT/mTOR pathway, the MAPK pathway, and the ERBB family of receptor tyrosine kinases were found in 91% of tumors. In line with these alterations, high mTOR signaling was frequently observed in OCCC, and mTORC1/2 inhibition was highly effective in OCCC cell lines and PDX models. These results indicate that especially targeting mTOR1/2 could be an effective therapeutic approach for patients with OCCC.

ARID1A was the most common mutated gene in our OCCC dataset, with a mutation rate of 46.6%. Overall, the mutation rates observed for ARID1A, PIK3CA, TP53, and KRAS matched earlier sequencing studies on smaller OCCC datasets (3, 4, 6, 8, 9, 24). Moreover, mutation incidences in PIK3CA, TP53, and KRAS correlated with the frequencies found in “pure OCCC” tumors as described by Friedlander and colleagues (22), indicating that our sample cohort also contains pure OCCC tumors. ARID1A and PIK3CA mutations have been shown as early-onset mutations in the development of OCCC (23, 25). Previous studies have shown that onset of oncogenesis regularly requires two or more mutational hits in a proto-oncogene or tumor-suppressor gene, as was demonstrated in a de novo OCCC mouse model (26). Indeed, in our tumor cohort, 94% of the PIK3CA- and ARID1A-mutant tumors contained at least one additional mutation or CNA in PI3K/AKT/mTOR, MAPK, or DNA repair pathway genes or the ERBB family of receptor tyrosine kinases, which was stage independent. Considering earlier reports and our observed OCCC mutation pattern, we hypothesize that a mutation in ARID1A or PIK3CA is accompanied by either a CNA or mutation in one of the aforementioned pathways to promote onset of OCCC. A limitation of the present study can be the imbalance in tumor samples and matched controls. Accordingly, we could designate a restricted number of mutations as somatic. Not all clinical data were available; therefore, we may have underestimated the relation between kinome alteration status and clinicopathologic characteristics or clinical outcome.

Genes mutated at low frequency (1%–10%) are generally difficult to designate as oncogenic drivers. Therefore, we used a large OCCC dataset and binomial testing of observed kinome variants. Eleven genes were found to be significantly mutated relative to the background mutation rate, and some of these had not previously been described in OCCC. Validation of variants in these low-mutation-frequency genes across multiple cancer types using the COSMIC database enabled us to predict AKT1, PIK3R1, FBXW7, ERBB3, ATM, CHEK2, and MYO3A as novel drivers of OCCC (27). Functional effects of AKT1 and PIK3R1 mutations in these domains were described in previous research. For instance, mutations in the AKT1 PH domain were shown to constitutively activate AKT1 (18). Likewise, mutations in the PIK3R1 inter-SH1–SH2 domain resulted in loss of PIK3R1 function, thus generating PIK3CA hyperactivation (19). These studies support the assumption that AKT1 and PIK3R1 are oncogenic drivers in OCCC. Furthermore, AKT1, PIK3R1, FBXW7, and ERBB3 all contribute to mTOR signaling, whereas ATM, CHEK2, and FBXW7 are implicated in DNA repair signaling (19, 21, 28). MYO3A mutations are found infrequently in cancer, but are still implicated in resistance to trastuzumab in breast cancer (29).

In the present study, kinome sequencing and SNP array analysis revealed substantial percentages of EGFR and ERBB2 mutations or amplifications in OCCC tumors. EGFR amplifications (associated with endometriosis, a precursor of OCCC) and ERBB2 amplifications were previously demonstrated in OCCC by comparative genomic hybridization (30, 31). Furthermore, sensitivity toward the clinically available inhibitors gefitinib and lapatinib was shown in OCCC cell lines harboring EGFR or ERBB2 alterations and may be explored further in the clinic (32, 33). In previous studies, EGFR-targeting therapy such as gefitinib or erlotinib and EGFR-ERBB2–targeting therapy using lapatinib did not provide clinical benefit in pretreated patients with ovarian cancer (34–36), but these studies did not focus on patients with OCCC. In addition, alterations in DNA repair pathway genes were identified in 82% of tumors, suggesting opportunities to target DNA repair cascades in OCCC in future research. Unfortunately, PARP1/2 inhibition by olaparib was not effective in our OCCC cell line panel, including BRCA1-mutant cell lines. Interestingly, inhibition of PARP1/2 by the PARP-trapping agent talazoparib demonstrated efficacy in a subset of OCCC cell lines (37).

Mutations in mTOR have been detected in several tumor types including HGS ovarian cancer (38). Although mTOR mutations were not found in OCCC tumors, alterations in upstream pathways that promote mTORC1/2 activation were found in 91% of these tumors. In line with these results, abundant high p-S6 expression was observed in OCCC tumors compared with HGS ovarian cancer tumors. Moreover, OCCC cell lines demonstrated nanomolar range sensitivity toward the mTORC1/2 inhibitor AZD8055 and higher sensitivity compared with HGS ovarian cancer cell lines. To our knowledge, we were the first to exploit the use of PDX models of OCCC. In vivo administration of AZD8055 resulted in a significant growth-inhibitory effect in three OCCC PDX models, consisting of a PIK3CA mutant, an ARID1A mutant, and a PIK3CA and ARID1A wild-type model. These models reflect the genetic makeup of OCCC. Hence, we propose mTORC1/2 inhibition as a future treatment strategy in OCCC that should be explored with urgency. Additional inhibition of ARID1A synthetic lethal targets in combination with mTORC1/2 inhibition can be utilized in ARID1A-mutant OCCC tumors (14, 15).

In the clinic, inhibition of mTORC1 with temsirolimus added to standard chemotherapy did not improve overall survival in newly diagnosed patients with stage 3 to 4 OCCC as compared with historical controls (39). However, by blocking both mTORC1 and mTORC2, reactivation of PI3K and MAPK signaling via mTORC2 can be prevented in OCCC. This may contribute to a better response in patients with OCCC. AZD8055 and another mTORC1/2 inhibitor, OSI-027, were tested in non-OCCC cancer types. Disappointingly, antitumor efficacy was only observed above MTD, and both drugs are no longer in clinical development (40, 41). Based on our results, it is conceivable that tumor responses in patients with OCCC can be attained at or below MTD with AZD8055 or OSI-027. A new generation of mTORC1/2 inhibitors, such as MLN0128 (Sapanisertib), is now being evaluated in multiple phase II clinical trials (NCT02724020 and NCT02725268). Given the large variation in mutations in specific genes of the PI3K/AKT/mTOR and MAPK pathway between patients and the high variation in sensitivity toward PI3K and MEK inhibition in our OCCC cell lines, PI3K or MEK monotherapy tumor responses in patients with OCCC will presumably be limited. MLN0128 is now being combined with a PI3Kα inhibitor (MLN1117) in an ongoing phase II trial in advanced endometrial cancer (NCT02725268). Drugs targeting PI3K as well as mTORC1/2, for example, dactolisib, DS-7423, and XL765, showed high toxicity in the clinical setting, despite promising efficacy in in vitro cancer models, including OCCC (16, 42–45). Based on our mutational analysis in patients with OCCC, drug sensitivity screens, and molecular analyses in OCCC cell lines, a strategy that combines inhibition of mTORC1/2 with PI3K or MEK will have the highest likelihood to improve clinical benefit for the majority of patients with OCCC. However, the use of single target inhibitors of mTORC1/2, PI3K, and MEK at suboptimal concentrations is probably crucial to maximize efficacy and minimize toxicities that were observed with PI3K–mTORC1/2 dual inhibitors. Overall, our findings set the stage to explore mTORC1/2-targeting phase II clinical trials in OCCC in the future.

E.I. Braicu is a consultant/advisory board member for Clovis, Roche Pharma, and Takeda. No potential conflicts of interest were disclosed by the other authors.

Conception and design: J.J. Caumanns, K. Berns, G.B.A. Wisman, M.J. Birrer, I. Vergote, R. Bernards, A.G.J. van der Zee, S. de Jong

Development of methodology: J.J. Caumanns, R.S.N. Fehrmann, H. Itamochi, M.J. Birrer, S. de Jong

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J.J. Caumanns, K. Berns, H. Klip, E.M. Hijmans, A.M.C. Gennissen, D. Weening, H.B. Salvesen, I. Vergote, E. van Nieuwenhuysen, J. Brenton, E.I. Braicu, J. Kupryjanczyk, B. Spiewankiewicz

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): J.J. Caumanns, G.B.A. Wisman, R.S.N. Fehrmann, T. Tomar, E.M. Hijmans, A.M.C. Gennissen, E.W. Duiker, R.J.C. Kluin, A.K.L. Reyners, I. Vergote, L. Mittempergher, R. Bernards, S. de Jong

Writing, review, and/or revision of the manuscript: J.J. Caumanns, K. Berns, G.B.A. Wisman, R.S.N. Fehrmann, E.W. Duiker, A.K.L. Reyners, M.J. Birrer, I. Vergote, E. van Nieuwenhuysen, J. Brenton, E.I. Braicu, J. Kupryjanczyk, A.G.J. van der Zee, S. de Jong

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): J.J. Caumanns, H. Klip, G.J. Meersma, E.M. Hijmans, A.M.C. Gennissen

Study supervision: G.B.A. Wisman, R. Bernards, A.G.J. van der Zee, S. de Jong

We acknowledge Drs. Els Berns, Yasushi Saga, and Vijayalakshmi Shridhar for their kind contribution of OCCC cell lines. We acknowledge the University of Cambridge, National Institute for Health Research Cambridge Biomedical Research Centre, Cambridge Experimental Cancer Medicine Centre, and Hutchison Whampoa Limited for their support. The authors thank Jolanda A.L. Visser, Neeltje M. Kooi, Gerda de Vries, and Fernanda X. Rosas Plaza for help with the PDX models. In addition, we acknowledge Anna Piskorz, Mercedes Jimenez-Linan, and Karen Hosking. G.B.A. Wisman and S. de Jong are members of the EurOPDX Consortium.

This research was supported by grants from the Dutch Cancer Society (KWF, RUG 2010-4833, 2011-5231, and 2012-5477), the Cock-Hadders foundation, and Cancer Research UK (A15601).

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.