Increased FOXJ1 protein expression is associated with improved overall survival in high-grade serous ovarian carcinoma: an Ovarian Tumor Tissue Analysis Consortium Study

Background Recently, we showed a >60% difference in 5-year survival for patients with tubo-ovarian high-grade serous carcinoma (HGSC) when stratified by a 101-gene mRNA expression prognostic signature. Given the varied patient outcomes, this study aimed to translate prognostic mRNA markers into protein expression assays by immunohistochemistry and validate their survival association in HGSC. Methods Two prognostic genes, FOXJ1 and GMNN, were selected based on high-quality antibodies, correlation with protein expression and variation in immunohistochemical scores in a preliminary cohort (n = 134 and n = 80, respectively). Six thousand four hundred and thirty-four (FOXJ1) and 5470 (GMNN) formalin-fixed, paraffin-embedded ovarian neoplasms (4634 and 4185 HGSC, respectively) represented on tissue microarrays from the Ovarian Tumor Tissue Analysis consortium underwent immunohistochemical staining and scoring, then univariate and multivariate survival analysis. Results Consistent with mRNA, FOXJ1 protein expression exhibited a linear, increasing association with improved overall survival in HGSC patients. Women with >50% expression had the most favourable outcomes (HR = 0.78, 95% CI 0.67–0.91, p < 0.0001). GMNN protein expression was not significantly associated with overall HSGC patient survival. However, HGSCs with >35% GMNN expression showed a trend for better outcomes, though this was not significant. Conclusion We provide foundational evidence for the prognostic value of FOXJ1 in HGSC, validating the prior mRNA-based prognostic association by immunohistochemistry.

countries [1]. Based on a prognostic signature of 101 genes, we have recently shown a dramatic difference (>60%) in median 5-year overall survival when stratified by gene expression score quintiles [2]. While the median survival time for the most favourable quintile reached 9.5 years, it was only 2.3 years in the least favourable quintile. This effect size outperforms any individual validated prognostic marker to date and further supports a remarkable biological heterogeneity within HGSC [3][4][5][6].
Historically, a blanket approach was applied to the treatment of women diagnosed with HGSC, though it is now undergoing a dramatic shift due to high response rates to poly (ADP-ribose) polymerase (PARP) inhibitors in BRCA1/2-deficient cases, allowing them to be added as maintenance to the first line of therapy [7,8]. However, this targeted therapy does not show the same efficacy across all patients with HGSC, and resistance may occur [9]. Therefore, other biomarkers related to prognosis may help to further delineate the biology of this aggressive disease. Candidate biomarkers that were discovered by mRNA expression profiling could be validated by immunohistochemistry (IHC). IHC is utilised routinely in clinical diagnostics across the globe, thus supporting its use as a tool for fast-tracked, statistically powered biological validation, particularly if large existing cohorts of tissue microarrays (TMAs) are also employed [10,11].
The purpose of this study was to translate previously identified prognostic mRNA targets into clinically applicable IHC assays and to validate their prognostic association at the protein level in HGSC and the other main OC histotypes.

Study design and participants
Sections of archival ovarian tumour tissue microarrays (TMA) were submitted by 21 studies from the Ovarian Tumor Tissue Analysis (OTTA) consortium (Supplementary Table 2) for analysis of GMNN and FOXJ1. Ethics approval was obtained for this project (University of New South Wales Human Research Ethics Advisory panel, #HC16299) and for participants through written informed consent or an Institutional Review Board ethics approval waiver (Supplementary Table 2). A retrospective cohort of 8798 and 7662 patient samples represented by individual or replicate TMA cores with a primary diagnosis of OC was assembled for FOXJ1 and GMNN staining, respectively. Contributing studies provided clinical covariates, including patient age at diagnosis, the time from diagnosis to OTTA study admission, histotype, stage of disease (FIGO (International Federation of Gynecology and Obstetrics) Stage I-IV, or localised/regional/distant), final vital status and overall survival (OS), defined as the time from diagnosis to death or last follow-up.

Immunohistochemistry of FOXJ1 and GMNN
Positive and negative on-slide controls for FOXJ1 and GMNN protein expression were used to assess stain specificity (Supplementary Table 3). Healthy pre-menopausal and peri-menopausal fallopian tube samples (n = 1 case each), were also examined through haematoxylin and eosin staining, and IHC for FOXJ1 and GMNN ( Supplementary Fig. 1).
IHC scoring was performed by two observers (AW and EYK). Observers were blinded to the clinicopathological data; one observer scored cases for FOXJ1 using a Nikon Eclipse 80i microscope (Nikon Inc., Chicago, IL, USA) at ×200 and another observer scored cases for GMNN using ImageScope v12.2.2.5015. Nuclear staining in each tumour core was semi-quantified by scores representing the percentage of immunopositive tumour cells in the core, in 5% intervals. Where cores were absent or tumour cells represented <25% of the core, no score was given. The maximum score was taken in cases with multiple cores on the TMAs. Intratumoural heterogeneity was assessed by a comparison of duplicate cores in n = 3401 (FOXJ1) and n = 3057 (GMNN) HGSC cases. Interobserver variability in FOXJ1 and GMNN expression scoring was evaluated in n = 221 and n = 311 cases, respectively.
FOXJ1 and GMNN expression score stratification IHC scores were visualised through frequency distributions of FOXJ1and GMNN scores in OC, and by histotype. Score stratification models simplified IHC scores. Thresholds used in stratification were selected to preserve the shape of distribution in the frequency histogram of HGSC scores. As FOXJ1 and GMNN expression differed in the distribution of their IHC scores, different score stratification models were applied to each data set, being FOXJ1 0%, 5%, 10-15%, 20-45% and 50-100%, and GMNN 0%, 5%, 10-15%, 20-25%, 30% and 35-100%. Score stratification was required to assess the association between expression and OS.

Correlation analysis of mRNA and protein expression
The correlation between mRNA expression data [2] and IHC protein expression score was assessed. To ensure score stratification models would not alter this relationship, the correlation between mRNA and stratified FOXJ1 and GMNN protein was also assessed.

FOXJ1 and GMNN protein expression in only chemotherapynaive samples
Sensitivity univariate and multivariate survival analyses in HGSC patient where treatment by primary debulking surgery was confirmed (n = 4440 (FOXJ1) and n = 4009 (GMNN)) were performed. This was to ensure the effects of neoadjuvant chemotherapy (NACT) did not impact the trends detected in the main survival analysis. Paired tissue samples taken before and after NACT were also assessed for FOXJ1 and GMNN protein expression (n = 23 and n = 21 respectively) to determine whether there were changes in expression in the tissue following exposure to chemotherapy.

Statistical analysis
The correlation between mRNA and protein expression was measured for each marker by Pearson correlation coefficients; the explainable variance was examined through the coefficient of determination. Clinical variables in FOXJ1 and GMNN cohorts were assessed through Chi-square testing of proportions. Kruskal-Wallis testing assessed IHC score differences by histotype. Squared-weighted Cohen's kappa statistics and frequency count matrices were used to assess intratumoural heterogeneity across duplicate TMA cores from the same case. Squared-weighted Cohen's kappa statistics were used to estimate concordance between multiple observers. A twotailed Wilcoxon-Pratt matched-pairs signed rank test was used to assess protein expression differences in matched pairs of HGSC tumour samples taken before and after NACT.
Histotype-specific univariate and multivariate analysis of the survival associations of stratified FOXJ1 and GMNN expression were performed, with death by any cause being the primary end point. Right censoring at 10 years was applied to account for deaths not related to OC. Left truncation mitigated against survival bias introduced by time between diagnosis and enrolment into a study. Kaplan-Meier curves with corresponding log-rank testing and calculation of year-specific survival, were used to compare prognosis between strata of the score stratification model. Multivariate Cox proportional hazards regression modelling estimated hazard ratios (HRs) and corresponding 95% confidence intervals (CIs). Models were adjusted for age, stage and strata of score and stratified by OTTA study site. Assumptions of proportional hazards were tested. In cases of non-proportional covariate hazards or large sample sizes, scaled Schoenfeld residuals were plotted to assess the violation. Adjusted HRs for each strata of the score model were visualised through forest plots, and used to identify a linear relationship between expression and the hazard of death. All statistical analyses were carried out using RStudio v1.1.463 or GraphPad Prism v7.02. R packages "survival" (version 3.3-1) and "survminer" (version 0.4.9) were employed in univariate and multivariate survival analysis. Statistical significance was defined by p < 0.05.

RESULTS
Selection of FOXJ1 and GMNN for analysis by IHC mRNA expression analysis by NanoString was previously performed on 3769 HGSC cases from the Ovarian Tumor Tissue Analysis (OTTA) consortium and 276 genes significantly associated with patient survival were identified [2]. The 90 genes of the greatest significance were examined for suitability for IHC analysis (Supplementary Table 4). We focussed on nuclear markers (avoiding stromal or immune markers due to challenges in interpretation of staining) with high-quality antibodies suitable for IHC in formalin-fixed paraffin-embedded tissues. Five genes met these criteria: GMNN, SNRPA1, FEN1, HIST1H2BD and FOXJ1. The previously identified HRs for one standard deviation change in gene expression, 95% CIs and adjusted q-values for these genes were: GMNN: 0.85 (95% CI 0.82-0.89) q = 1.89 × 10 −11 , SNRPA1: 0.87 (95% CI 0.83-0.91) q = 1.84×10 −10 , FEN1: 0.89 (95% CI 0.85-0.92) q = 1.44 × 10 −7 , HIST1H2BD: 0.89 (95% CI 0.85-0.93) q = 1.44 × 10 −7 , FOXJ1: 0.90 (95% CI 0.86-0.94) q = 2.56 × 10 −6 [2]. Protein expression of these markers was assessed by IHC in a small cohort of 80-134 cases. The distribution of IHC scores were examined; FOXJ1 and GMNN expression showed the greatest variation in scores ( Supplementary Fig. 2). Expression of these markers was positively skewed with median expression scores of 5 and 20%, respectively. The correlation between mRNA expression data from NanoString analysis and protein expression was determined ( Supplementary Fig. 2). A significant correlation between mRNA and protein expression was identified for FEN1, FOXJ1 and GMNN ( Supplementary Fig. 3). The levels of mRNA were able to explain some variance in FOXJ1 r 2 = 0.283 and GMNN r 2 = 0.23 protein expression. Both FOXJ1 and GMNN were selected for further analysis as they showed prognostic significance at the mRNA level, demonstrated a significant correlation between mRNA and protein, and had variation in IHC scores, suggesting they could be suitable for prognostic stratification of patients.
Protein expression of GMNN and FOXJ1 across OC histotypes, and the relationship with mRNA expression Of the 8798 (FOXJ1) and 7662 (GMNN) cases, 2364 and 2192 cases respectively were excluded due to diagnosis with a histotype other than HGSC, LGSC, MC, EC or CCC, no survival data, uninterpretable or insufficient tumour tissue (<25% of the tumour core), overlapping cases between TMAs or withdrawn consent. The clinicopathological parameters by histotype for 6434 FOXJ1 and 5470 GMNN cases that met the inclusion criteria were evaluated (Table 1 and Supplementary Table 5). Subsets of HGSC cases had 80% power to detect a HR of 1.3, with a type-1 error rate of 0.001. OC histotypes had significantly different FOXJ1 and GMNN expression patterns (Fig. 1). FOXJ1 protein expression varied more than GMNN expression across histotypes with greater interquartile ranges of expression scores in all histotypes (apart from MC). Intratumoural heterogeneity was assessed by weighted Cohen's kappa statistics in duplicate HGSC TMA cores when FOXJ1 and GMNN expression were given as 5% intervals (Cohen's kappa 0.41 and 0.40 respectively) and stratified IHC scores (Cohen's kappa 0.74 and 0.72 respectively; Supplementary Fig. 4). 3111 FOXJ1 and 2900 GMNN HGSC cases also had mRNA expression data collected through NanoString nCounter analysis [2]. The correlation between mRNA and protein expression was preserved in a HGSC subset of the final cohort to undergo survival analysis (Fig. 1). Stratification of FOXJ1 and GMNN protein expression did not appear to alter this correlation (Fig. 1).
Association of stratified GMNN and FOXJ1 protein expression and OS, by histotype Stratified FOXJ1 protein expression was significantly associated with OS in univariate (p < 0.0001) and multivariate (p = 0.0002) analysis of HGSC survival ( Fig. 2 and Table 2). Higher FOXJ1 expression was significantly associated with a higher probability of survival, relative to cases with absent or lower expression (Table 2). Prognosis was most favourable in tumours with >50% of cells expressing FOXJ1 (HR 0.78, 95% CI 0.67-0.91, p < 0.05), with the 5-year survival being >10% higher than those without expression.
Univariate and multivariate analysis of stratified FOXJ1 protein expression in CCC indicated a significant association with improved survival (p = 0.0096 and p = 0.02, respectively;  Table 2). Stratified GMNN protein expression was not significantly associated with OS in HGSC (p = 0.14), though tumours with >35% GMNN expression appeared to be associated with an improved OS (Fig. 2). This trend was not supported by multivariate survival analysis adjusting for age and stage, although the overall p value for GMNN was significant (p = 0.001) ( Table 3). Univariate analysis of stratified GMNN expression in EC and CCC indicated a significant association with survival ( Supplementary Fig. 6), where higher expression in both cases conferred improved outcomes. However, adjusted analysis indicated that stratified GMNN expression was not significantly associated with EC or CCC patient survival (p = 0.09 and p = 0.65, respectively; Table 3). Adjusted analysis of GMNN expression in the two LGSC cases with >30% expression in tumours had significantly poorer outcomes (HR 8.28, 95% CI 1.36-50.43, p < 0.05; Table 3).
The proportional hazards assumption of Cox regression modelling was not violated by age, stage or OTTA site, in any histotypes, for both markers. To account for age and stage artefacts in these results, induced by the large sample sizes of HGSC-specific analyses, Schoenfeld residuals were plotted and assessed ( Supplementary Fig. 7). No significant deviations from the line of best fit were detected. The univariate and multivariate survival analysis results were replicated in a sensitivity analysis of a subset of the HGSC cohort, using only known chemotherapy naïve tissue, from patients where primary debulking surgery was the confirmed first line of therapy ( Supplementary Fig. 8 and Supplementary Table 6).

DISCUSSION
Our study validates the linear, increasing prognostic association of FOXJ1 protein expression in HGSC in a large, international cohort. While our analysis of FOXJ1 exemplifies the power of IHC for rapid biological validation of prognostic biomarkers that were originally discovered as mRNA candidates, the lack of translation of other markers also illustrates the challenges.
As FOXJ1 protein expression increased, the probability of survival in HGSC increased, implying expression has a linear association with outcome. Patients with tumours with >50% of cells expressing FOXJ1 had the longest survival. Sui and colleagues have already suggested that high FOXJ1 expression is associated with favourable OC prognosis [12]. The present analysis uses a statistically powered, histotype-specific approach to differentiate that FOXJ1 is only associated with prognosis in HGSC and does not indicate the prognosis of other histotypes, perhaps except for CCC. High expressing CCC were associated with favourable survival as were cases with lower expression, contrasting the observed linear association with survival in HGSC. Sui and colleagues proposed that the stem cell-related transcription factor NANOG is a negative regulator of FOXJ1-mediated OC migration and invasion [12]. This suggests that NANOG high/FOXJ1 low are more akin to cancer stem cells, which may be more likely chemoresistant and thereby less susceptible to chemotherapy, leading to poor prognosis. However, we were unable to address the response to chemotherapy in the current data set due to lack of data, e.g. the Response Evaluation Criteria in Solid Tumors (RECIST) data. Yet this generates an interesting hypothesis for future studies.
The biological significance and prognostic association of FOXJ1 in cancer is varied [13,14]. Survival associations consistent with our HGSC results have been observed in gastric cancer, as well as ependymomas and choroid plexus tumours, where high expression of FOXJ1 is thought to be a marker of better tumour differentiation and a more favourable prognosis [15,16]. It may be appealing to conclude that FOXJ1 is a marker for welldifferentiated HGSCs. FOXJ1's characteristic role is in motile ciliogenesis, and we confirmed FOXJ1 expression in ciliated, but not secretory cells of the fallopian tube. However, this would somewhat contradict the prevailing concept that HGSC arises from the secretory cells, as opposed to ciliated cells, of the fallopian tube via precursor serous tubal intraepithelial carcinoma (STIC) lesions and diffuse PAX8 staining, a marker of secretory cells, is present in virtually all HGSCs, supporting the STIC hypothesis [17][18][19][20][21]. As motile cilia are neither a feature of STIC nor HGSC, the expression of FOXJ1 does not suggest that HGSC arise from terminally differentiated ciliated cells of the fallopian tube or that a cell lineage switch occurs [22]. Instead, the expression of the ciliated marker FOXJ1 in HGSC can be considered as a marker of enigmatic differentiation, which does not result in morphologically apparent cilia and is associated with better outcome; a phenomenon that has been previously suggested for endometrial cancers [23]. "Aberrant" FOXJ1 expression in HGSC may reflect epigenetic reprogramming of secretory cell-derived tumour cells or a unique differentiation state of the cell of origin. A recent study proposed that HGSC are composed of a mixture of cancer cell states inherited from its cell of origin in the fallopian tube including a ciliated state [24]. An enigmatic ciliated differentiation is further supported by expression of other ciliated markers (such as Ezrin) in OC confirming that FOXJ1 expression is not an isolated, aberrantly expressed marker of ciliogenesis, but represents a more general differentiation programme [25,26]. In addition to FOXJ1's potential in HGSC prognostication, future studies may use FOXJ1 to gain further insights into HGSC development.
Contrary to results in FOXJ1, the univariate survival analysis with GMNN expression indicated that there was no significant prognostic association in HGSC. Contrastingly, multivariate survival analysis did indicate that GMNN expression was significantly associated with HGSC survival, although there were no observable trends in HRs across protein expression groups that suggested this association was linear. In a threshold effect, individuals with >35% of tumour tissue expressing GMNN did demonstrate a trend towards a more favourable prognosis, although this was not supported by the 95%CI in the multivariate analysis. As GMNN's significant mRNA association with survival did not translate into protein expression, GMNN is not currently suitable for prognostic  [27,28]. In the final cohort, the correlation between mRNA and GMNN protein expression was weaker, relative to FOXJ1 expression. While this difference was slight, it may have contributed to the result. Dissimilar mRNA and protein abundances may be reflective of the translation rate, its modulation, delays in protein synthesis and transport, and the modulation of the proteins' turnover-rate [29][30][31]. As a surrogate marker of proliferation, GMNN's expression is highly regulated during the cell cycle [32]. Protein translation and degradation may be uncoupled from mRNA transcription and stability, and may explain the lack of association with survival for GMNN protein levels. Another explanation might be that the FOXJ1 IHC assay was superior in quality and therefore easier to interpret as shown by the better interobserver reproducibility. Perhaps another reason that GMNN did not translate into a suitable prognostic IHC marker is the less variable expression within HGSC, when compared to FOXJ1. While prognostic significance was not shown in LGSC, multivariate analysis indicated that individuals with >35% of tumour tissue expressing GMNN were associated with a more unfavourable prognosis. However, as this group only included two individuals this should only be considered an interesting observation highlighting the challenges of biomarker studies in rare histotypes even in a consortium type approach. No associations were seen in the other histotypes. GMNN expression did however differ significantly in samples taken before and after NACT in small cohort of HGSC cases. This difference can potentially be attributed to the role of GMNN in cell division alongside the impact of chemotherapy on this process [33]. However, trends that would suggest GMNN expression increased or decreased following NACT were not observed. Furthermore, the potential clinical relevance of this result is limited by the small sample size, which would not completely capture disease heterogeneity. HGSC cases known to be chemotherapy naïve underwent univariate and multivariate survival analysis to confirm the prognostic associations revealed in this study.
The potential for FOXJ1 and GMNN expression stratification into broader score-groupings was observed in our HGSC cohort. As 0% and 5%, as well as 10%-15% and 20%-45% exhibited consistently similar survival associations in the cohort, analysis of FOXJ1 expression by 0%-5%, 10-45% and 50%-100% IHC score groupings may prove more-appropriate for clinical assessment. Similarly, the trend observed in tumours with high GMNN expression suggested that stratification into 0%-35% and 40%-100% may clarify the relationship with survival, though the lack of significant results in this study, does render the value of further GMNN investigation dubious.
We acknowledge that the translational potential of our findings is limited by heterogeneity within HGSC. Future work may take important molecular alterations within HGSC, such BRCA1/2 mutation or homologous repair deficiency status, into account. Further, we recognise that the immediate translational potential of FOXJ1 as an individual prognostic biomarker is limited given the observed survival increase of~10% in a small subset (less than 10%) of HGSC patients. Yet we believe that due to robust prognostic associations FOXJ1 should be evaluated in prognostic multimarker models.
Several variables can interfere with the translation of mRNA targets into IHC biomarkers. Our marker selection was severely limited by the number of genes with available high-quality monoclonal antibodies, which were necessary as they ensured increased stain reproducibility across FFPE TMA sections. We avoided proteins of the microenvironment and focussed on those with nuclear expression as they are often the most robust markers for established clinical tests. When selecting markers, we observed only moderate correlations between mRNA and protein expression for the five initially selected markers. While gene expression or protein stability can influence these correlations, we believe that a main factor is the use of conventional IHC, which is not quantitative in nature [34][35][36][37]. Modern IHC detection systems greatly amplify the signal leading to a quick saturation. This might explain why some of the initial markers showed ubiquitous expression by IHC in comparison to variable mRNA expression levels. In addition, markers were assessed on TMAs with an intrinsic limitation in assessing for intratumoural heterogeneity. However, our assessment of intratumoural heterogeneity in cases with duplicate TMA cores demonstrated strong agreement in FOXJ1 and GMNN protein expression scores. While this may not capture all variation in protein expression across each tumour, it does support that the scores used in our analysis were representative of the tumour profile.
Overall, our study has provided foundational evidence that the prognostic mRNA signature of FOXJ1 can be translated into an IHC biomarker, to stratify prognosis in HGSC. Survival improved with stratified, increasing FOXJ1 protein expression. It indicated this relationship could be further stratified, to improve clinical utility, which is intriguing regarding its unexpected expression in HGSC. We also demonstrate the challenges of validating prognostic IHC markers. The careful study design and power has ensured these results provide much-needed insights into the likely outcomes of HGSC patients.

DATA AVAILABILITY
Participants of this study did not agree to their data being shared publicly; accordingly, the data used in this research will not be made available.