Estimation of PTENexpression in high-grade serous ovarian carcinoma is strongly influenced by stromal content

We evaluated the stromal content of 216 HGSOC samples from TCGA in a total of 302 images using a computational framework validated through scoring by an independent observer (Jonckheere–Terpstra test for trend P=0.001) (Figure 1A and Additional file 1: Figure S1). The automated stromal scores were highly correlated with the expression of genes from a published stromal gene signature (Figure 1B) [22]. ACTA2 ranked 17 in the top correlated stromal genes and was therefore selected for subsequent analysis on the basis of its known stromal-specific expression (Figure 1C) [23].

High ACTA2 expression in the TCGA samples was directly correlated with PTEN expression and was never associated with low PTEN values, suggesting that in the majority of samples it was stromal PTEN expression that was being measured (Figure 1D).

Differential gene analysis comparing the upper and the lower quartiles of PTEN expression showed enrichment for stromal genes in tumours with high PTEN (Gene Set Enrichment Analysis (GSEA) Enrichment Score = 0.5). However, performing the analysis on samples with low ACTA2 content (the first quartile) showed a more random distribution of stromal genes (GSEA ES=0.1), suggesting this subset is less influenced by stromal content (Figure 1E,F). The wider distribution of PTEN expression in quartile one of ACTA2 expression also supports the hypothesis of tumour PTEN loss being more prevalent than previously estimated (Figure 1D).

PTENloss is prevalent and has prognostic value in high-grade serous ovarian carcinoma

To test the predictions that reduced PTEN expression could be a frequent event, we developed methods to quantify tumour-specific PTEN expression using a semi-quantitative immunofluorescence (IF) procedure. We applied this to tissue microarrays constructed from the population-based Study of Epidemiology and Risk Factors in Cancer Heredity (SEARCH) cohort (N=245 HGSOCs from 516 ovarian cancer samples; Table 1) [3]. For HGSOC, PTEN expression was variable and showed a range of intensities, from negative (23%;N=49), weak positive (32%;N=68) to strongly positive fluorescence (23%;N=49). Heterogeneous fluorescence was also observed in 48 samples, 22% of the cases (Figure 2A).

Reduced PTEN fluorescence (including negative, weak or heterogeneous expression) was associated with significantly worse survival for HGSOC compared with positive fluorescence, independent of study site, age, stage and grade (hazard ratio 1.8, 95% confidence interval (CI) 1.0 to 3.0, P=0.03; Figure 2B, Additional file 2: Table S1).

We examined for interactions between BRCA1 and PTEN loss of expression as PTEN loss is a frequent initiating event in BRCA1-associated breast tumours and is associated with basal-like breast cancer [8]. All patients with a deleterious germ-line BRCA1 mutation had HGSOC tumours and were marginally more likely to have negative or weak PTEN staining (Fisher’s exact test P=0.06, N=9/10 and Additional file 2: Table S1).

PTENis frequently deleted in high-grade serous ovarian carcinoma

TP53 is mutated in more than 95% of HGSOC cases [2],[4] and has been previously implicated in controlling PTEN transcription [24]. We tested for TP53 effects on PTEN expression by introducing a bacterial artificial chromosome transgene containing the entire human TP53 locus into the TP53-null cell line SKOV3. PTEN expression was independent of TP53 complementation for both wild-type and mutated (R175H, R273H) transgenes (Figure 3A). Comparison of promoter methylation and expression of PTEN in the TCGA data set showed infrequent methylation of the PTEN promoter region in low-PTEN expressing cases (Figure 3B).

By contrast, loss of a single PTEN allele was common in HGSOC (36%; N=174) in addition to the previously described homozygous deletion occurring in 6% of tumours [4] (Figure 3C). PTEN gene expression was significantly lower when there was loss of at least one allele (Figure 3D; t-test, P≪0.001). To assess whether protein expression of PTEN was associated with copy number, we applied our IHC staining classification to a subgroup of 51 samples from the TCGA cohort that has been recently stained for PTEN [14]. A large number of positive samples were reclassified as weak positive (Figure 3E,F and Additional file 2: Table S3). These tumours had similar levels of PTEN mRNA to those with heterogeneous staining, and significantly lower levels than tumours with positive staining (Wilcoxon test, P=0.002; Figure 3G), emphasising the importance of differentiating intensities in PTEN staining. Negative staining was strongly correlated with homozygous deletion, and weak or heterogeneous staining with hemizygous loss; positive staining was associated with no chromosomal loss (Fisher’s exact test, P=0.001; Figure 3H).

Androgen receptor expression is associated with PTENexpression

We analysed PTEN differentially expressed genes for TCGA samples with low ACTA2 content to mitigate the effect of stromal contamination. AR was one of the top differentially expressed genes (Figure 1E), and this was confirmed using an orthogonal method, csSAM, which takes into account stromal content from H&E images (Additional file 4).

Protein–protein interaction data from TCGA available through cBioPortal [25] showed that the highly ranked phosphorylated proteins in the lower quartile of PTEN expression (defined as PTEN RNA-sequencing expression, z score<−0.5) included AKT1, AKT2 and AKT3, which reflects activation of the PI3K pathway. Additionally, another highly ranked protein expressed in this subgroup was AR. A direct link between AR and PTEN was suggested by overlaying genomic information, including copy number and gene expression, on a known protein interaction network (Figure 4A). Moreover, AR expression has also been associated with PTEN expression in prostate cancer [26]. Therefore, we hypothesised that AR expression was prognostically significant. Using the TCGA RNA-sequencing data, we found that low AR expression was associated with shorter overall survival (hazard ratio 1.5, 95% CI 1.1 to 2.1, P=0.02; Figure 4B).

These results were validated experimentally in 216 samples from the SEARCH cohort. It was found that 43%, 25% and 32% of HGSOC expressed high levels of AR (≥50% of tumour cells), low levels of AR (<50% of tumour cells) or no AR, respectively (Figure 4C). In a multivariate analysis, we found a similar prognostic effect for expression in these samples (hazard ratio 1.7, 95% CI 1.1 to 2.4, P=0.01), despite there being no strong association between AR and PTEN (chi-squared test P=0.63; Figure 4D,E).

Differentiated and proliferative expression subtypes are associated with high and low PTENexpression

TCGA subdivided HGSOC into four subgroups based on gene expression profiles, differentiated, proliferative, mesenchymal and immunoreactive, in which the latter two were enriched for stromal cells and leukocytes (Figure 5A) [4],[18].

PTEN loss is associated with activation of the PI3K pathway and consequently with proliferation. Therefore, we hypothesised that PTEN expression could be associated with the remaining differentiated and proliferative subgroups. By focusing on these two subgroups and categorising the raw AR and PTEN expression into tertiles, higher PTEN expression was associated with the differentiated subgroup (Figure 5B, chi-squared test P=0.02), whilst lower PTEN expression was associated with the proliferative subgroup.

Overview of data sets used

Correcting gene expression profiles using image analysis

Scoring by eye for overall image quality (based on discolouration, folding over of the mounted section and completeness of the section) was performed on 312 slides, of which 46 were discarded owing to poor quality, leaving 266 slides from 194 patients. Based on quality, 302 H&E slides from 216 patient samples in the TCGA database were selected, of which an observer (FCM) manually scored 266 for stromal content. Images were segmented by first applying an entropy filter to remove the background from an image, followed by colour deconvolution according to Ruifroks’ method (Additional file 5) [39]. The haematoxylin channel was subtracted from the eosin channel, leaving a raw stromal signal. Otsu’s thresholding and smoothing was then performed to estimate a stromal fraction [40]. These values along with stromal gene scores were used to predict the stromal content in the remaining TCGA samples. Correlation between automated and manual scoring was performed using the Jonckheere–Terpstra test for trend. Gene-expression-based validation of the method was performed by generating a stromal gene list [22] and performing univariate Pearson’s correlation testing between stromal quantification and expression.

Differential gene expression analysis

PTEN expression was categorised into quartiles and the top and bottom quartiles were used for differential expression analysis, which was performed using either the limma package [41] or an orthogonal method, csSAM [42], on the complete TCGA set (N=489) and in the subset of low ACTA2 tumours (N=122). Differentially expressed genes were selected after correction for the false discovery rate for multiple testing and using a cut-off P<0.05. The top 50 differentially expressed genes were selected for further unsupervised hierarchical clustering and visualisation using the made4 made4 library [43]. Unsupervised hierarchical clustering was performed using the Euclidean distance metric and Ward’s method.

Statistical tests and survival analysis

All statistical analysis were performed using R [44]. All the R code necessary to replicate the data analysis is available in Additional file 6. Survival analysis was performed using a Cox proportional hazards model. Since some patients died just after diagnosis and were not included in the SEARCH study, left truncation was included in the analysis of this cohort, which means that we took into account both the time from diagnosis to entry into the study and the time from entry into the study to censoring or death.

PTEN immunostaining, scanning and scoring of SEARCH and NOT samples

IHC was carried out using tissue microarrays obtained from Formalin Fixed, Paraffin-Embedded (FFPE) tissues and primary antibodies for PTEN and AR (Cell Signaling, Danvers, MA, USA; PTEN – Clone 138G6; AR – Clone D6F11). The staining protocol for PTEN clone 138G6 was previously described by FCM [8] and validated for specificity and sensitivity using Pten knockout mice and xenograft tissue from PTEN-positive and -negative human cancer cell lines (see Supplementary Figure 1 in [8]). Heat-induced antigen retrieval was carried out in 10 mmol ⁻¹ citric acid (pH 6.0) in a pressure cooker at 120°C for 10 min. Sections were incubated with PTEN or AR antibodies overnight, diluted 1:100 in 5% goat serum, followed by incubation with anti-rabbit biotinylated secondary antibody for 1 h and peroxidase-conjugated avidin-biotin complexes (Elite ABC; Vector Laboratories, Burlingame, CA, USA). Formed immunocomplexes were visualised using diaminobenzidine (DAKO, Glostrup, Denmark, EU) and slides were counterstained with haematoxylin. Sections were rinsed in PBS between each step.

IF was carried out by adding a tyramide signal amplification step (Perkin-Elmer) after the secondary antibody, followed by incubation with Alexa Fluor 647-conjugated streptavidin (Invitrogen). Nuclei were counterstained with 4^′,6-diamidino-2-phenylindole (DAPI) (Invitrogen).

IF and IHC samples were stored at −20°C and room temperature (RT), respectively, for at least 48 h before image analysis. For IHC, an Ariol scanning system (Leica, Wetzlar, Germany, EU) was used to obtain the digital images. For IF, images were acquired with a SP5 Leica Confocal Microscope, ×40 plan objective, and analysed by Leica software (Leica Application Suite, Advanced Fluorescence 2.2.0). Scoring was performed according to intensity (using stromal cells as internal positive controls) and percentage of stained cells by two independent observers (FCM and MJL). We subdivided tumours as negative (no staining in any tumour cell), weak positive (all tumour cells weakly stained compared to stromal cells), positive (all tumour and stromal cells equally stained) or heterogeneous (combination of positive and negative/weak staining) staining.

Western-blot PTEN quantification in relation to TP53status

Western blot analysis was performed for extracts from SKOV3 ovarian cancer cell lines (ATCC; HTB-77) and derivatives obtained by Bacterial artificial chromosome (BAC) lipofectamine transfection. A BAC modification kit (Gene Bridges, Heidelberg, Germany, EU; catalogue number K002) was used to obtain the BAC clones (empty control and with R175H or R273H mutations) from the original BAC containing wild-type TP53 (CTD-3049A20; Invitrogen, catalogue number 96012). Whole-cell extracts were collected after scraping cells in protein lysis buffer containing 50 mmoll ⁻¹ Tris-HCl (pH 7.4), 150 mmoll ⁻¹ NaCl, 5 mmoll ⁻¹ ethylenediaminetetraacetic acid (EDTA), 50 mmoll ⁻¹ NaF, 0.5% NP40, one Complete™ Mini and EDTA free tablet per 50 ml. Protein concentrations were determined using Bio-Rad protein assay kit (Hercules, CA, USA). Equal concentrations were resolved by sodium dodecyl sulphate polyacrylamide gel electrophoresis (SDS-PAGE), transferred to Immobilon-fluorescence PVDF membrane (Millipore, Bedford, MA, USA), blocked and probed with anti-human TP53 (clone DO-1, Santa Cruz, Dallas, Texas, USA; 1:2,000; overnight incubation at 4°C), anti-human PTEN (clone 138G6, Cell Signaling; 1:1,000; overnight incubation at 4°C) and anti GAPDH (Cell Signaling; 1:5,000; overnight incubation at 4°C) antibodies. IRDye800-conjugated anti-rabbit immunoglobulin G (IgG) and IRDye700-conjugated anti-mouse IgG (Li-Cor, Lincoln, NE, USA; 1:5,000 and 1:10,000 dilutions; incubated at RT for 1 h) were used as secondary antibodies. Signal intensities were analysed by using the Odyssey infrared image system (Li-Cor, Lincoln, NE, USA). Two replicates of the experiment were performed and similar results were obtained.