Association of p16 expression with prognosis varies across ovarian carcinoma histotypes: an Ovarian Tumor Tissue Analysis consortium study

Abstract We aimed to validate the prognostic association of p16 expression in ovarian high‐grade serous carcinomas (HGSC) and to explore it in other ovarian carcinoma histotypes. p16 protein expression was assessed by clinical‐grade immunohistochemistry in 6525 ovarian carcinomas including 4334 HGSC using tissue microarrays from 24 studies participating in the Ovarian Tumor Tissue Analysis consortium. p16 expression patterns were interpreted as abnormal (either overexpression referred to as block expression or absence) or normal (heterogeneous). CDKN2A (which encodes p16) mRNA expression was also analyzed in a subset (n = 2280) mostly representing HGSC (n = 2010). Association of p16 expression with overall survival (OS) was determined within histotypes as was CDKN2A expression for HGSC only. p16 block expression was most frequent in HGSC (56%) but neither protein nor mRNA expression was associated with OS. However, relative to heterogeneous expression, block expression was associated with shorter OS in endometriosis‐associated carcinomas, clear cell [hazard ratio (HR): 2.02, 95% confidence (CI) 1.47–2.77, p < 0.001] and endometrioid (HR: 1.88, 95% CI 1.30–2.75, p = 0.004), while absence was associated with shorter OS in low‐grade serous carcinomas (HR: 2.95, 95% CI 1.61–5.38, p = 0.001). Absence was most frequent in mucinous carcinoma (50%), and was not associated with OS in this histotype. The prognostic value of p16 expression is histotype‐specific and pattern dependent. We provide definitive evidence against an association of p16 expression with survival in ovarian HGSC as previously suggested. Block expression of p16 in clear cell and endometrioid carcinoma should be further validated as a prognostic marker, and absence in low‐grade serous carcinoma justifies CDK4 inhibition.


Introduction
CDKN2A (cyclin-dependent kinase inhibitor 2A) is located on chromosome 9p21.3 and encodes two proteins, p16 and p14 ARF , that have different reading frames [1]. p14 ARF inhibits p53 function and p16 inhibits the CDK4/6 complex acting as a negative cell cycle regulator suppressing the transition from the Gap1 to DNA synthesis (G1/S) phase and arresting the cell cycle in the G1 phase [2]. Normal cells express variable amounts of p16 protein that can be detected by immunohistochemistry (IHC) in both nuclear and cytoplasmic localizations (heterogeneous p16 expression pattern) [3]. There are two abnormal p16 expression patterns: absent and overexpressed, the latter also referred to as block expression as recommended by the Lower Anogenital Squamous Terminology Standardization Project for HPV-Associated Lesions (LAST) [4]. In keeping with its role as a tumor suppressor, absence of p16 expression can occur due to various mechanisms including homozygous deletion, loss of function mutations, promoter hypermethylation and translational suppression [5]. In ovarian carcinoma, homozygous deletion of CDKN2A has been detected in only 3% of high-grade serous carcinomas (HGSC) [6], 15% of low-grade serous carcinomas (LGSC) [7], and in 30% of mucinous carcinomas (MC) [8]. In contrast, p16 block expression results from a variety of alterations in G1/S cell cycle transition as a compensatory effort to inhibit G1/S transition. p16 block expression is classically observed in human papillomavirus (HPV)-associated uterine cervical neoplasms, in which viral proteins (E7) inactivate pRB and promote G1/S transition [9,10]. IHC overexpression of p16 is routinely used in clinical diagnostics for identification of HPV-related neoplasms. Ovarian carcinomas are not associated with HPV infections, but alterations promoting G1/S transition are common, e.g. RB1, CCNE1, CCND1, or MYC [6].
Ovarian carcinoma is a biologically heterogeneous disease [11] composed of five main histotypes: HGSC, LGSC, clear cell carcinoma (CCC), endometrioid carcinoma (EC), and MC, which should be studied separately [12]. Older studies combining all histotypes showed that either overexpression or complete absence of p16 were associated with unfavorable outcomes [13][14][15]. Recently, histotype-specific studies also reported that normal heterogeneous p16 expression was significantly associated with longer progression-free and overall survival (OS) in two series of 334 and 115 women with HGSC [16,17]. Therefore, we hypothesized that heterogeneous p16 expression reflecting the normal G1/S transition status is associated with a favorable outcome in HGSC compared to absent or block expression reflecting abnormalities of the G1/S cell cycle checkpoint complex. The purpose of this study was to validate whether abnormal p16 expression is associated with an unfavorable OS in HGSC, and to explore prognostic associations in other histotypes using tissue microarrays (TMAs) from the Ovarian Tumor Tissue Analysis (OTTA) consortium [18,19].

Immunohistochemistry
The study investigators obtained tissue from 7492 patients with a diagnosis of primary ovarian, fallopian tube, or peritoneal carcinoma from 24 study sites (Supporting Information, Table S1). Most of these patients also participated in previous OTTA studies [18][19][20], and all studies received ethics board approval for tumor profiling. TMAs were constructed containing 1-6 cores of 0.6-1.0 mm in diameter from formalin-fixed paraffin embedded tissue representing tumor from each patient. p16 IHC was performed centrally at two institutions: Genetic Pathology Evaluation Centre, University of British Columbia, and Calgary Laboratory Services, University of Calgary, Canada. TMAs were stained in five batches with three different protocols (Table S2) using the same antibody (clone E6H4, CINtec, mtm laboratories). Three staining patterns were recorded: absent, heterogeneous and block ( Figure S1). Block expression was distinguished from heterogeneous staining by using the recommendation for p16 interpretation from LAST [4]; that is, block expression is characterized by diffuse staining of tumor cells in nuclear and/or cytoplasmic compartment with at least moderate intensity with virtually no negative tumor cell clusters. Interobserver reproducibility between two observers (PR and MK) was assessed for a subset of 120 cases. Seventeen studies were scored by PR and the remainder by MK. Cases represented by more than one core and discordant cores were consolidated as heterogeneous if any of a given case score was heterogeneous.

CDKN2A mRNA analysis
A subset of 2280 cases had CDKN2A mRNA expression data from NanoString n-counter analysis. RNA was extracted from 10 μm sections from formalin fixed paraffin embedded (FFPE) tissue blocks, which were macrodissected to avoid adjacent benign tissue but included tumor stroma using the Qiagen miRNeasy (Qiagen Inc. Toronto, Ontario, Canada) FFPE protocol and quantitated on a Nanodrop spectrophotometer (Thermo-Fisher Scientific, Waltham, MA, USA). After mixing 500 ng of total RNA per sample with a custom codeset (NanoString Technologies Inc, Seattle, WA, USA) and hybridization buffer (NanoString), hybridization was performed using a Tetrad 2 thermal cycler (Bio Rad Laboratories Inc, Hercules, CA, USA) for 16 or 20 h and then analyzed on a nCounter Digital Analyzer (NanoString). Data was normalized to housekeeping genes (RPL19, ACTB, PGK1, SDHA, and POLR1B) and pre-processed to a reference of 3 pooled ovarian cancer specimens as described previously [21]. We interrogated the cBioportal for associations of CDKN2A alterations with OS in HGSC from TCGA [22,23].

Statistical tests
Morphology-based histotype was derived from pathology reports with or without review of reports or slides (Table S1). Because some HGSC were mistakenly classified as other histotypes in the past, we used the highly specific WT1(+)/p53(mutant) IHC combination to reclassify those to HGSC [24]. We excluded 409 cases owing to diagnosis other than the five major histotypes, 393 cases being uninterpretable, 31 cases with a combination of absence and block staining, and 134 cases with missing clinical follow-up data. The final sample size was 6525 ( Table 1). The median time from diagnosis to enrollment was 0 days (interquartile range 0-182 days). Patients (n = 331) with missing 253 CDKN2A/p16 in ovarian carcinoma data for either age or time from diagnosis to enrolment were not part of the multivariate survival analysis.
Associations of p16 IHC expression and CDKN2A mRNA expression with demographic and clinical variables were examined using the chi-square test and Kruskal-Wallis test, as appropriate. We examined interobserver heterogeneity of p16 interpretation using Kappa coefficients. For individuals with multiple tumor cores, we examined intratumoral heterogeneity of p16 expression using percent discordance. We assessed correlations between (the ordinally scaled) p16 staining values and CDKN2A mRNA expression using Pearson correlation coefficients.
The primary end point for survival analysis was death from any cause. We chose right censoring of follow-up at 12 years post-diagnosis to mitigate against deaths from noncancer-related causes. Kaplan-Meier survival curves and corresponding log-rank tests were generated to visually assess associations of p16 expression with survival. Cox proportional hazards regression was used for multivariable assessment of hazard ratios (HRs). Models were adjusted for the following confounding factors: study site, age (continuous), FIGO stage (categorized into I/II, III/IV, and missing variable), and residual disease (categorized as absent, i.e. no residual disease, present, and missing). We used left truncation to account for the enrollment of prevalent cases in some studies. We tested whether histotype modified the association between p16 IHC expression and OS by fitting and testing corresponding interaction terms. We assessed the functional form of the association between CDKN2A mRNA expression levels and OS in HGSC using penalized Bsplines [25], adjusting for the same potential confounding variables as described above. All statistical tests were two-sided, and analyses were carried out using RStudio (Boston, MA, USA) or JMP 13.0.0 (SAS, Cary, NC, USA). This study adhered to the REMARK criteria [26].

Results
p16 protein and CDKN2A mRNA expression across histological types Table 1 shows the characteristics of the study sample. The histotype distribution is similar to population-based cohorts except for a slightly higher frequency of MC [27]. Inter-observer agreement for interpretation of p16 IHC was excellent (Cohen's kappa of 0.92). Of the 6525 women, 4046 (62.0%) had more than one interpretable tissue core and, of those, 12.9% had a discordant interpretation between cores. This moderate degree of intratumoral heterogeneity was not statistically different between histotypes (p = 0.11), ranging from 7.4% for LGSC to 16.3% for CCCs. As expected, the distribution of p16 expression categories was significantly different across histotypes (  Figure 1A).

Association of p16 protein and CDKN2A mRNA expression and OS in HGSC and LGSC
For HGSC patients, the Kaplan-Meier survival curve showed no difference in OS for the three p16 immunohistochemical expression patterns ( Figure 1B, p = 0.32), which was supported by HRs near 1.0 after controlling for the study, age, time to enrollment, stage and residual disease ( Table 2). For mRNA expression in HGSC, we used several different groupings (median split, tertiles) as well as different cut-offs for dichotomization (mean = −3.03, visual inspection to separate bimodal peaks = −3.7) but there was no association of CDKN2A mRNA levels with OS. Alternatively, we examined the functional form of the association between CDKN2A expression and OS in HGSC using penalized B splines [25]. Analyses revealed relatively flat HRs across the entire spectrum of mRNA values with a 95% confidence band that always included an HR of 1.0 ( Figure S3). By interrogating 489 HGSC from TCGA via the cBioPortal [22,23], 92 (19%) showed downregulation of CDKN2A, which was also not associated with survival (p = 0.27).
In contrast, for patients with LGSC, the 5-year survival rate was significantly lower in tumors with absent p16 expression ( Figure 1C, 43.4%, SE 10.7%) and in tumors with block p16 expression (42.2%, SE 14.7%) compared to heterogeneous expression (70.1%, SE 3.8%, p = 0.0005). This was also significant in multivariate analysis for the absence of p16 ( Table 2). The limited number of CDKN2A mRNA expression values for non-HGSC precluded us from examining associations within those histotypes.

Pooled association of p16 protein expression and OS, and tests of effect modification by histotype
In analyses that combined all five major histological subtypes, women whose tumors exhibited with block expression had poorer OS (HR 1.13, 95% CI 1.06-1.22) than those with absent or heterogeneous expression ( Figure S4). The associations between p16 expression and survival were strikingly different across histotypes (Cox regression test for interaction p = 1.4 × 10 −8 ).

Association of p16 protein expression with clinicopathological parameters within histotypes
In EC, a greater proportion of cases with p16 block expression were diagnosed at stage III/IV (33.8% compared to 14.5% for heterogeneous staining,  Table S3. Notably, there was no association for p16 expression status with BRCA1 or BRCA2 mutation status for the subset of HGSCs with available mutation data (n = 1370, p = 0.43).

Discussion
Our investigation showed that associations of p16 staining pattern with OS differ across ovarian carcinoma histotypes. Block p16 expression was significantly associated with shorter survival for endometriosis-related ovarian carcinomas. In contrast, absence of p16 expression predicted shorter survival in LGSC while no survival associations are observed for MC and HGSC.
In contrast to previous studies [16,17], we provide strong evidence against a clinically or biologically relevant survival association of p16 expression in HGSC. Our large sample size greatly reduces the potential of this lack of association being a false negative finding. Using our observed sample size and a two-sided test of hypothesis with Type I error rate of 0.05, we would have 80% power to detect a HR as low as 1.21 comparing absence and 1.10 comparing block to heterogeneous expression, respectively. This null conclusion is supported by a lack of association between OS and CDKN2A mRNA expression data, which correlated positively with p16 IHC-based protein expression. In keeping with the recommendations from the Institute of Medicine for validation of biomarker studies [29], we used the same antibody and the same scoring system as previous studies. We found excellent interobserver reproducibility regarding the IHC interpretation and similar frequencies of the three staining patterns in HGSC compared to previous studies; and it is, therefore, unlikely that technical or interpretational differences can explain differences in our results from those published previously. We think that the large sample size used in the current study compared to prior studies of HGSC avoided a false positive finding [16,17]. Since p16 block staining is a surrogate for various retinoblastoma pathway alterations, we speculate that different underlying alteration might explain the lack of outcome associations for p16. For example, prognostically opposing underlying alterations (e.g. favorable pRB loss versus unfavorable CCNE1 amplifications), which result in the same p16 block staining, may neutralize each other [30,31].
Exploring other histotypes, we demonstrate for the first time that block p16 expression is significantly associated with OS in both endometriosis-associated histotypes: EC and CCC. Overall, those histotypespecific differences would not have been revealed in a combined histotype analysis and corroborate that biomarker analyses should be done stratified by histotype [12]. Yet the subsets of p16 block expression that were significantly associated with unfavorable prognosis were small: 9 and 14% of EC and CCC, respectively. EC is usually associated with a favorable outcome and some patients do not require chemotherapy or could be considered for hormonal therapy if hormone receptors (ER, PGR) are expressed [18,32]. However, estrogen receptor positive Luminal B breast cancers with loss of pRB function and high p16 expression are unresponsive to hormonal therapy [33]. Our data suggest that further study of p16 as part of a biomarker panel that identifies EC with unfavorable prognosis would help triage patients to earlier aggressive therapy in the low stage setting, which may not be amenable to hormonal therapy.
We observed a similar negative association between block p16 expression and OS for CCC. Women diagnosed with CCC have a relatively unfavorable prognosis, in part because these tumors are chemotherapy-resistant and alternative therapeutic options are sparse [34]. Among those, radiation has been suggested for CCC [35] and perhaps p16 expression can be assessed for prediction of response to radiation as suggested from other cancer sites [3,36]. The survival associations specifically observed for p16 block staining in the two endometrioisisassociated histotypes somewhat serves as a cross-validation. Yet it does not preclude differences in the underlying mechanisms, e.g. CCNE1 copy number gain and overexpression have been reported for CCC but not EC [37]. As a limitation, we observed a moderate degree of intratumoral heterogeneity, which was highest in CCC, using TMAs in size akin to pretreatment omental core biopsies. However, CCCs are usually treated by upfront surgery and the p16 assessment on a full histological section should mitigate against intratumoral heterogeneity.
We also observed a significant association with OS in patients with LGSC. In contrast to the block staining in endometriosis-associated carcinomas, complete absence was associated with unfavorable outcome in LGSC in keeping with the tumor suppressor function. Although investigating the underlying mechanism of absence of p16 expression is beyond the scope of the current study, the 12% of LGSC showing complete absence of p16 by IHC is strikingly similar to the 15% frequency of the homozygous deletion of the CDKN2A locus reported by Hunter et al [7]. We have previously shown that progesterone receptor (PGR) expression is a favorable prognostic factor in LGSC [18]. Perhaps PGR and p16 status could help to stratify LGSC regarding prognosis [38]. Another consideration is the possible predictive utility of absent p16 expression for CDK4 inhibitors as suggested in clinical trials for breast and other cancers [39]. Konecny et al demonstrated that low p16 expression in pRB-proficient 257 CDKN2A/p16 in ovarian carcinoma tumors was correlated with in vitro response to CDK4 inhibitors [40]. Since other treatment options are limited for LGSC, this may represent an interesting option for recurrent LGSC, a disease often affecting younger women. There was a non-significant trend for the few LGSC with p16 block expression to have an unfavorable outcome but CKD4 inhibitors will be ineffective in tumors with p16 block expression because p16 block expression already indicates futile intrinsic CDK4 inhibition.
There was no significant prognostic association of p16 expression within MCs despite their having the highest frequency of complete absence (50%) across histotypes. This frequency is slightly higher than the 39% (n = 12/31) frequency of the combined corresponding molecular alterations (homozygous CDKN2A deletion or loss of functional mutations) reported by Ryland et al [8]. Absence of p16 expression was most frequently observed in low stage MC and we speculate that a portion of those may be from low transcriptional activity.
A strength of this study is the large sample size providing excellent power to examine differences in protein expression patterns for the most common histotype (HGSC) as well as reasonable power to discern differences within EC and CCC. The lower sample sizes for MC and LGSC reduced our ability to detect differences. We used a diagnostic biomarker panel of WT1/p53 to limit the number of misclassified HGSC into other histotype categories. Within the OTTA consortium, we had the opportunity to analyze protein and mRNA data. Survival analyses were adjusted for confounding factors such as age, stage and residual disease. As a limitation, some of the study sites did not collect detailed treatment data; as such adjustment for treatment was not feasible. However, cases were collected throughout an era of relatively homogeneous standard adjuvant therapy consisting of platinum-taxane chemotherapy before the introduction of neoadjuvant chemotherapy or PARP inhibitors.
This large-scale collaborative study did not validate p16 expression as a prognostic marker in HGSC. The frequent block expression as a surrogate for abnormal retinoblastoma pathway activation warrants further study of individual pathway members. The intriguing prognostic associations in endometriosis-associated EC and CCC make p16 a promising prognostic biomarker that requires further independent validation. The absence of p16 in a subset of LGSC calls for independent validation as a prognostic marker as well as investigation as a predictive marker for CDK4 inhibitors.

Author contributions statement
DGH and MK conceived the study design. PFR, CC, MSA, SMA, CW, AT and MK carried out experiments and interpreted results. RAV and MK analyzed the data. All authors collected data. PFR, RAV and MK wrote the first draft. All authors were involved in writing the paper and had final approval of the submitted and published versions.