c-MET/VEGFR-2 co-localisation impacts on survival following bevacizumab therapy in epithelial ovarian cancer: an exploratory biomarker study of the phase 3 ICON7 trial

All participants provided written informed consent for the International Collaboration of Ovarian Neoplasms 7 (ICON7) trial (ClinicalTrials.gov number, NCT00262847): an international, randomised, multi-centre, open-label, phase 3 trial investigating the use of bevacizumab maintenance therapy as part of the first-line therapy in women diagnosed with high-risk FIGO (1988) stage I–IIA (grade 3/poorly differentiated tumours or clear cell carcinoma) or FIGO (1988) IIB-IV epithelial ovarian, primary peritoneal or fallopian tube cancer [5]. Progression-free survival (PFS) was defined as the time interval from the date of randomisation to the date of disease progression (defined clinically or radiologically, but not using CA125 criteria) or death. Overall survival (OS) was defined as the time interval from the date of randomisation to the date of death from any cause. All demographic and survival data were provided by the clinical trial centre co-ordinating ICON7 (Medical Research Council Clinical Trials Unit, University College London). As part of written informed consent, ICON7 participants could optionally agree to donate whole blood samples and pre-chemotherapy tumour tissue for translational research. Data presented here is from all patients who agreed to donate tumour tissue and/or whole blood samples. No case selection has been carried out.

Tissue microarrays (TMAs) from EOC tumour tissue sections underwent immunofluorescent staining using the BenchMark ULTRA system according to the RUO DISCOVERY universal staining procedure (Ventana Medical Systems). TMA sections were deparaffinised at 69°C for 8 min per cycle. Antigen retrieval occurred by pre-treatment with Cell Conditioning Solution 2 (CC2) (Ventana Medical Systems) and incubation at 95 °C for 4, 8, 16, 24 and 32 min. TMA sections were then blocked by incubating with DISCOVERY inhibitor (Ventana Medical Systems) for 8 min. Sections were warmed to 37 °C and then incubated with the following primary antibodies: VEGF receptor 2 (55B11) rabbit monoclonal antibody (dilution 1:600; Cell Signalling Technology) and c-MET (1G7NB) mouse monoclonal antibody (dilution 1:300; Novus Biologicals) for 60 min. Sections were then incubated with the corresponding HRP-conjugated secondary antibodies for 8 min. The Tyramide Signal Amplification (PerkinElmer) was used on all slides to improve visualisation following incubation with the secondary antibody. Tyramide labelled with Cy3 (to visualise VEGFR-2, dilution 1:50; PerkinElmer) or FITC (to visualise c-MET, dilution 1:50; PerkinElmer) diluted in 1X Plus Amplification Diluent (PerkinElmer) were used.

Following the first staining, the second staining on the same section was performed using anti-human collagen type IV (dilution 1:300; Merck Millipore) to identify the stroma in each core. Sections were pre-treated with CC2 at 95 °C for 4 min and then 8 min. Slides were then cooled to 37 °C before the second primary antibody was applied manually, followed by incubation for 60 min. Slides were then incubated with a corresponding HRP-conjugated secondary antibody for 16 min followed by incubation with the Cy5 fluorophore-conjugated tyramide (dilution 1:50) diluted in 1X Plus Amplification Diluent for 8 min. Slides were removed from the instrument and washed twice in EZ Prep buffer (Ventana Medical Systems). SlowFade Gold Antifade Mountant with 4′,6-diamidino-2-phenylindole (DAPI) was added, and slides were mounted with coverslips.

Leica Aperio VERSA-200 digital pathology scanner was used to visualise immunofluorescence staining. Images were acquired at × 20 magnification using the Andor iXon 888 camera (Additional File 1: Fig. S1). Immunofluorescence images from TMA sections were analysed by molecular pathology image analysis using Definiens Tissue Studio 2.7. Each TMA core was segmented into tumour and stroma by machine learning using collagen IV staining to identify stroma. VEGFR-2 staining was used to detect vessels, and to calculate number of vessels, vessels density (number of vessels per tumour area), median vessel size and VEGFR-2 expression (mean of stain intensity) in the tumour component of each TMA core. c-MET/VEGFR-2 co-localisation was quantified using c-MET total intensity in VEGFR-2 positive vessel. The co-localisation values were not standardised by tumour area, vessel number and sizes because this information was included separately in data analysis.

Peripheral blood samples were collected in K2EDTA blood tubes. Germline DNA was extracted from the precipitated leukocyte cell fraction. Genotyping was performed at the VIB Center for Cancer Biology (Leuven, Belgium) using the Sequenom MassARRAY® system (iPLEX GOLD) as reported previously [15]. Single nucleotide polymorphisms (SNPs) in VEGF-A and VEGFR-2 were analysed. A detailed description of the way in which SNPs were selected in these genes has been published previously [15, 16].

In the ICON7 trial, bevacizumab maintenance therapy showed an PFS advantage of hazard ratio (HR) 0.81 compared to placebo [5]. A group of patients were however identified whom appeared to be disadvantaged by bevacizumab [8, 9]. The HR between those patients whom benefitted versus those that did not was assumed to be 0.56, a 30% reduction from the original HR of 0.81. Based on a two-sided Cox proportional hazard regression analysis, it was estimated that a total number of 124 patients will be needed to achieve a significance level of 0.05 and a power of 0.8, assuming data censorship was less than 25%. Including a 15% contingency cohort, a minimum patient size of 143 would therefore be required. In total, samples from all 178 patients who gave consent to donate tumour tissue were employed in this study.

Variables from three types of data were analysed in this study: (1) clinical data, (2) immunofluorescence biomarker data including VEGFR-2 and c-MET expression and (3) genotyping biomarker data for SNPs in VEGF-A and VEGFR-2. The analysis started from investigating the associations between immunofluorescence variables and treatment outcomes (data types 1 and 2), followed by identifying SNPs that were associated with selected immunofluorescence biomarkers in matched patients (data types 2 and 3) and finally exploring the association between the identified SNPs and treatment outcomes (data types 1 and 3).

The relationship between variables in clinical, immunofluorescence and genotyping data were assessed using one-way ANOVA for continuous variables or chi-squared tests for categorical variables.

The pre-treatment prognostic and predictive significance of each variable were assessed by including the biomarker as a sole covariate (univariable) in a proportional hazards model for PFS or OS. Here, prognostic significance refers to whether there is a significant association between a biomarker and survival independent of treatment arms, while predictive significance refers to if the association between a biomarker and survival are significantly different depending on treatment arms. Variables were included in continuous form where possible and were subject to transformation, such as log₂ or dichotomisation by median. Assumption of proportionality was verified based on Schoenfeld residuals [17]. Martingale residuals from each marker specific analysis was examined for evidence of nonlinearity in the biomarker-hazard relationship [18]. Candidate biomarkers with P-values ≤ 0.05 in the univariable analysis were selected for subsequent multivariable proportional hazard regression analysis. A backward stepwise method was applied to identify the optimum subset of variables that associated with PFS or OS. Interactions between prognostic biomarkers of interest and treatment arms were included to seek predictive biomarkers that predicted benefit from bevacizumab. To address potential confounding factors, clinical variables that are significantly associated with immunofluorescence biomarkers are included in multivariable models regardless of their association with patient survival. Their interaction with immunofluorescence biomarkers were also interrogated.

Genotyping data from VEGF-A and VEGFR-2 were introduced to explore potential biological mechanisms underpinning the immunofluorescence data. The SNPs were filtered by their association with immunofluorescence biomarkers and those that strongly associated were investigated for their prognostic and predictive significance with respect to PFS/OS using the proportional hazard model described above. Analyses were carried out in accordance with the REMARK guidelines [19] and were implemented using R 3.6.0 [20].

To test the clinical significance of c-MET/VEGFR-2 co-localisation in samples from 178 patients diagnosed with EOC, a TMA-based immunofluorescence staining method was developed for c-MET and VEGFR-2. The statistics of key immunofluorescence biomarkers are summarised in Additional File 2: Table S1. The clinical data from these 178 cases was similar to the original ICON7 population with a slightly higher proportion of clear cell carcinoma cases in the immunofluorescence biomarker study, although the difference was not significant (Fisher’s exact test, P = 0.07) (Table 1). In this group of 178 patients, those with FIGO stage III/IV disease demonstrated significantly higher vessel density (P = 0.005), number of vessels (P = 0.004), median vessel size (P = 0.006) and VEGFR-2 expression (P = 0.003); consistent with the concept that angiogenesis is associated with tumour growth and metastasis (Additional File 2: Table S1).

The association between vascular tissue markers of angiogenesis and higher FIGO stage provided confidence in the validity of the immunofluorescence biomarkers and sample set, supporting further evaluation of the relationship between immunofluorescence biomarkers and survival outcomes. As a first step, a univariable analysis of clinical prognostic factors and vascular tissue biomarkers was performed (Additional File 2: Table S2). Both FIGO stage (III/IV vs. I/II) and amount of residual disease (≤ 1 cm vs. > 1 cm) following debulking surgery were of prognostic significance, in keeping with other ovarian cancer datasets (Additional File 2: Table S2) [6, 7]. In addition, number of vessels (P = 0.015) and VEGFR-2 expression (P = 0.027) inversely associated with PFS regardless of treatment (Additional File 2: Table S2). Importantly, in this univariable analysis, c-MET/VEGFR-2 co-localisation predicted significantly worse OS (P = 0.001) in patients treated with bevacizumab (Additional File 2: Table S2).

We next carried out a multivariable analysis to interrogate whether immunofluorescence biomarkers could predict treatment benefit of bevacizumab, but not placebo, with established prognostic clinical factors. Consistent with other ovarian cancer studies, FIGO stage (III/IV vs. I/II; hazard ratio [HR] 2.384, 95% confidence interval [95%CI] 1.181–4.815, P = 0.015) and amount of residual disease (≤ 1 cm vs. > 1 cm; HR 0.243, 95%CI 0.147–0.399, P < 0.001) were significant prognostic factors for PFS (Table 2). For OS, the amount of residual disease was also significant (≤ 1 cm vs. > 1 cm; HR 0.397, 95%CI 0.228–0.694, P = 0.001; Table 2). Multivariable analysis showed that greater c-MET/VEGFR-2 co-localisation predicted worse OS in patients treated with bevacizumab after adjusting for FIGO stage and debulking surgery outcome (HR 1.034, 95%CI 1.010–1.059, P = 0.006) (Table 2). Furthermore, by dichotomising patients according to the median expression of c-MET/VEGFR-2 into c-MET/VEGFR-2^LOW (≤ median) and c-MET/VEGFR-2^HIGH (> median) groups, data showed that patients with c-MET/VEGFR-2^HIGH tumours treated with bevacizumab had significantly reduced OS compared to c-MET/VEGFR-2^LOW tumours (39.3 vs. > 60 months; multivariable HR 2.00, 95%CI 1.08–3.72, P = 0.03; Fig. 1A) whereas in the placebo group, there was no significant difference (multivariable HR 1.26, 95%CI 0.67–2.39, P = 0.47; Fig. 1B).

Four hundred and forty-nine patients underwent SNP genotyping: 216 in the control group and 233 in the bevacizumab group (Table 1). Thirty-five VEGF-A and VEGFR-2 SNPs were investigated to understand the potential genetic mechanisms underpinning the variation in c-MET/VEGFR-2 co-localisation (Additional File 2: Table S3). Focusing on the 142 matched patients who underwent both immunofluorescence studies and genotyping, three SNPs were identified that were strongly associated with high co-localisation of c-MET and VEGFR-2. These included VEGF-A rs3025033 (G variant; chi-squared test, P = 0.138), VEGFR-2 rs1870377 (A variant; chi-squared test, P = 0.151) and VEGFR-2 rs2305945 (G variant; chi-squared test, P = 0.047) (Additional File 2: Table S3). The prognostic and predictive significance of these SNPs were evaluated, with univariable and multivariable results presented in Additional File 2: Table S3 and Table 3, respectively. Of these three SNPs, VEGFR-2 rs2305945 independently predicted PFS in the bevacizumab-treated cohort after adjusting for known prognostic clinical factors (multivariable HR 1.592, 95%CI 1.110–2.284, P = 0.012) (Table 3). Patients with the VEGFR-2 rs2305945 G/G variant who were treated with bevacizumab had shorter PFS than those with G/T or T/T variants (18.3 vs. 23.0 months; HR 0.74, 95%CI 0.53–1.03, P = 0.07; Fig. 2A). In contrast, there was moderately longer PFS in patients with the VEGFR-2 rs2305945 G/G variant who were treated with chemotherapy alone (21.2 [G/G variant] vs. 15.6 months [G/T or T/T variants]; HR 1.14, 95% 0.79–1.63, P = 0.48; Fig. 2B). The frequency of each variant in the three key SNPs are shown in Additional File 2: Table S4. The Kaplan-Meier curves of PFS for VEGF-A rs3025033 and VEGFR-2 rs1870377 SNPs are shown in Additional File 2: Fig. S2.

The two multivariable analyses show that c-MET/VEGFR-2 co-localisation (178 patients) and VEGFR-2 rs2305945 (449 patients) are independent predictive factors for bevacizumab-associated outcomes, supporting the hypothesis that there are a group of patients whose outcome is adversely affected by treating with bevacizumab. Furthermore, the VEGFR-2 rs2305945 G/G variant is associated with increasing c-MET/VEGFR-2 co-localisation (P = 0.047; Additional File 2: Table S3) and each biomarker predicts worse PFS (VEGFR-2 rs2305945: HR 1.592, P = 0.012; Table 3) and OS (c-MET/VEGFR-2 co-localisation: HR 1.034, P = 0.006; Table 2) in patients treated with bevacizumab, suggesting a possible biological and/or clinical interaction (Additional File 1: Fig. S3). An exploratory analysis on the relationship between c-MET/VEGFR-2 co-localisation and the VEGFR-2 rs2305945 SNP was carried out by investigating their interaction in one multivariable model (Additional File 2: Table S5). It was observed that patients had worse PFS and OS if they had high c-MET/VEGFR-2 co-localisation and the VEGFR-2 rs2305945 G/G variant, although the difference was not significant (P = 0.071 for PFS and P = 0.272 for OS; Additional File 2: Table S5).