The histone chaperone HJURP is a new independent prognostic marker for luminal A breast carcinoma

Abstract

Background

Breast cancer is a heterogeneous disease with different molecular subtypes that have varying responses to therapy. An ongoing challenge in breast cancer research is to distinguish high-risk patients from good prognosis patients. This is particularly difficult in the low-grade, ER-positive luminal A tumors, where robust diagnostic tools to aid clinical treatment decisions are lacking. Recent data implicating chromatin regulators in cancer initiation and progression offers a promising avenue to develop new tools to help guide clinical decisions.

Methods

Here we exploit a published transcriptome dataset and an independent validation cohort to correlate the mRNA expression of selected chromatin regulators with respect to the four intrinsic breast cancer molecular subtypes. We then perform univariate and multivariate analyses to compare the prognostic value of a panel of chromatin regulators to Ki67, a currently utilized proliferation marker.

Results

Unsupervised hierarchical clustering revealed a gene cluster containing several histone chaperones and histone variants highly-expressed in the proliferative subtypes (basal-like, HER2-positive, luminal B) but not in the luminal A subtype. Several chromatin regulators, including the histone chaperones CAF-1 (subunits p150 and p60), ASF1b, and HJURP, and the centromeric histone variant CENP-A, associated with local and metastatic relapse and poor patient outcome. Importantly, we find that HJURP can discriminate favorable and unfavorable outcome within the luminal A subtype, outperforming the currently utilized proliferation marker Ki67, as an independent prognostic marker for luminal A patients.

Conclusions

The integration of chromatin regulators as clinical biomarkers, in particular the histone chaperone HJURP, will help guide patient substratification and treatment options for low-risk luminal A breast carcinoma patients.

Abbreviations
• AFA
• Adaptive Focused Acoustics
• ASF1
• anti-silencing function 1
• BCSS
• breast cancer specific survival
• BLC
• basal-like cancer
• CAF-1
• chromatin assembly factor 1
• CENP-A
• centromere protein A
• CR
• chromatin regulators
• DFI
• disease-free interval
• DFS
• disease-free survival
• DMFS
• distant metastasis-free survival
• EMA
• easy microarray data analysis
• ELISA
• enzyme-linked immunosorbent assays
• ER
• estrogen receptor
• ESR1
• estrogen receptor gene
• EZH2
• enhancer of zeste homolog 2
• HJURP
• holliday junction recognition protein
• IHC
• immunohistochemistry
• MCM
• minichromosome maintenance
• MFI
• metastasis-free interval
• PARP
• poly(ADP-ribose) polymerase
• PGR
• progesterone receptor gene
• RPLP0
• ribosomal protein large P0
• RPPA
• reverse phase protein arrays
• TCGA
• The Cancer Genome Atlas

1 Introduction

Breast cancer is a heterogeneous disease with varying responses to therapy. While 60% of early stage breast cancer patients receive some type of chemotherapy, less than half clinically benefit from it (Schmidt et al., 2009; Weigelt and Reis-Filho, 2010). To help guide treatment plans, clinicians currently rely on four primary markers to classify tumors into four intrinsic molecular subtypes; basal-like (BLC), Erbb2 (or HER2), luminal A, and luminal B (Figure 1A) (Perou et al., 2000; Sorlie et al., 2001, 2003). These markers have enabled an improved overall breast cancer survival by identifying high-risk patients, often estrogen receptor (ER) negative, likely to respond to adjuvant therapy (O'Shaughnessy et al., 2011; Rouzier et al., 2005). However, in poor prognosis ER-positive patients, only 2–4% generally benefit from chemotherapy (Berry et al., 2006). This is key given that ER+ breast cancers constitute approximately 70% of all breast cancers (Mullan and Millikan, 2007). The proliferation-associated antigen Ki67 (herein referred to as MKI67 when discussing gene expression) (Gerdes et al., 1983) is currently used to distinguish the ER+ luminal A and luminal B tumors, however, despite widescale use, its application remains controversial (Maisonneuve et al., 2014; Polley et al., 2013). Further, while transcriptomic tests, including Mammaprint, Oncotype DX, and PAM50 (Chia et al., 2012; Cobleigh et al., 2005; Correa Geyer and Reis-Filho, 2009; Nielsen et al., 2014; Paik et al., 2004; van 't Veer et al., 2002; van de Vijver et al., 2002), use molecular profiling to provide additional prognostic information, their actual clinical value is still debated (Patani et al., 2013). Thus, distinguishing luminal B from luminal A tumors remains a critical challenge in breast cancer management and new markers are necessary to better guide clinical decisions (Harbeck et al., 2014).

Hierarchical clustering of 1127 breast cancer samples according to the expression of select chromatin regulators. A) Summary of features for defined molecular subtypes of breast cancer based on (Mullan and Millikan, 2007; Reis-Filho and Pusztai, 2011; Schnitt, 2010). B) Tumor specimens were divided into the four molecular subtypes (see methods). The cluster dendrogram (horizontal axis) shows the four subtypes of tumors colored as: basal-like, red; HER2+, green; luminal B, purple; and luminal A, light blue. The dendrogram of the 58 classification probe sets corresponding to the chromatin regulators is shown on the vertical axis. Each column represents a single patient, and each row represents a single gene. Green squares, transcript levels below the median; black squares, transcript levels equal to the median; red squares, transcript levels greater than the median. Color saturation reflects the magnitude of the ratio relative to the median for each set of samples (see scale, top right). C) Enlargement of a gene cluster representing a putative CR “subtype gene signature”.

It is well established that both genetic and epigenetic alterations contribute to tumorigenesis (Baylin and Jones, 2011). Recent work has highlighted how a misregulation of chromatin regulators, including histone variants (Filipescu et al., 2013; Vardabasso et al., 2014), histone chaperones (Gurard-Levin et al., 2014), histone-modifying enzymes (Gurard-Levin and Almouzni, 2014) and effector proteins, and chromatin remodelers, (Avvakumov et al., 2011; Polo and Almouzni, 2006) can participate in cancer initiation and progression (Elsheikh et al., 2009; Seligson et al., 2005). In one example, the expression of the histone chaperone holliday junction recognition protein (HJURP), responsible for depositing the histone variant centromere protein A (CENP-A, also known as CenH3) at the centromere (Dunleavy et al., 2009; Foltz et al., 2009), is upregulated in lung and breast cancer (Hu et al., 2010; Kato et al., 2007), and CENP-A is overexpressed in some of the most aggressive cancers (Gu et al., 2014; Lacoste et al., 2014; Li et al., 2011b; Qiu et al., 2013; Tomonaga et al., 2003; Wu et al., 2012). Additionally our previous work showed that the histone chaperone anti-silencing function 1b (ASF1b), but not its paralog ASF1a, is upregulated in many tumors, including breast (Abascal et al., 2013; Corpet et al., 2011), and has prognostic value for metastasis in breast cancer (Corpet et al., 2011). This evidence prompts an examination of how chromatin regulators may offer additional valuable information in the clinical setting.

Here we explore first the ability of a selection of chromatin regulators to classify breast cancer molecular subtypes, and second we examine their prognostic value independent of clinicopathological variables. For this we exploited a published transcriptome dataset (1127 patients) (Reyal et al., 2008), and we verified our results using an independent cohort of 71 patients from the Institut Curie. We find that not only do selected chromatin regulators differentiate luminal A from luminal B tumors, but the histone chaperone HJURP can also distinguish good and poor prognosis patients within the luminal A subtype, outperforming MKI67. Our data highlight the clinical value of chromatin regulators and support their integration in the clinic to improve cancer management.

2 Materials and methods 2.1 Selection of chromatin regulators of interest and microarray data set

We selected a subset of 54 chromatin regulators that includes histone chaperones, histone variants, histone modifying enzymes, and effector proteins (Table S1). In addition, we also included the estrogen (ESR1), progesterone (PGR), and ERBB2 receptors, and MKI67. We used a previously published and comprehensive transcriptome data set (Reyal et al., 2008) that compiled six independent breast cancer data sets, hybridized to HG-U133A Affymetrix arrays, and their associated clinical data, which were processed and normalized as previously described (Reyal et al., 2008). We classified patients into basal-like, HER2+, luminal B, and luminal A based on the mRNA expression levels from the estrogen receptor (ESR1, probe set rs205225_at) and ERBB2 (probe set rs216836_at) genes, in conjunction with the genome grade index as previously defined (Sotiriou et al., 2003).

2.2 Probe set selection and hierarchical clustering of microarray data

We verified the quality of each probe set in the NetAffx analysis center (http://www.affymetrix.com/estore/), choosing probe sets with annotation grade ‘A’; then, we verified these probes using the UCSC genome browser (http://genome.ucsc.edu/cgi-bin/hgGateway). Lastly, we applied the JetSet algorithm to select a single ‘optimal’ microarray probe set to represent each gene (Li et al., 2011a) (Table S1). We performed exploratory analyses to identify similar sample profiles using the R package EMA (Easy Microarray data Analysis) (Servant et al., 2010). We used Pearson correlation and ward criteria for unsupervised hierarchical clustering and visualized the output by a heat map that contained all genes arranged according to similar gene expression patterns across patient samples.

2.3 Statistical analysis of microarray data

We compared the mRNA levels of each gene per molecular subtype by a variance analysis (ANOVA) and performed pairwise comparisons by the Wilcoxon test, to test for significant associations between gene expression levels and breast cancer subtypes. Then, we carried out a survival analysis to investigate the impact of gene expression on disease outcome. Depending on the information available (Reyal et al., 2008) patient survival was defined either as distant metastasis-free survival (DMFS) or as breast cancer specific survival (BCSS). DMFS is defined as the time (in months) between surgery and the diagnosis of the first distant metastasis or the patient's death for patients who did not relapse. BCSS is defined similarly but considers only death due to the initial breast cancer. For the analyses, the expression level of each gene was transformed into a categorical variable by performing a median split of the distribution. Then, DMFS and BCSS curves were derived by the Kaplan–Meier method and compared by a log-rank test. Associations between gene expression and clinical outcome were determined by both univariate and multivariate Cox regression analyses, using a forward stepwise method and a p < 0.05 inclusion value. Corrections for multiple testing were done by the Benjamini–Hochberg test. The clinical variables included in the model were tumor size, lymph node, Elston-Ellis grade, estrogen receptor status, and molecular subtype.

2.4 Breast tumor samples for independent validation set

Tumor samples were from patients with breast carcinomas classified as lymph node-negative (N0), metastasis-free (M0), non-palpable (T0) or small (T1-T2) that were treated by primary conservative surgery at the Institut Curie (Table S2). A total of 71 patients diagnosed in 1996, gave informed consent to use their biological samples and clinical data for research purposes. The patients received tumor excision with radiotherapy. The median follow-up time was 153 months (range: 6–177 months). Recurrence-free and surviving patients were monitored periodically until the day of their last appointment. At the time of this analysis, 24% of the patients had died, 11% more developed loco-regional cancer recurrences, and 3.7% developed metastasis. The 71 tumors were classified at the Institut Curie Hospital as basal-like, HER2+, luminal B, or luminal A, where luminal B tumors were grade II/III with Ki67 expression ≥20% and luminal A tumors were grade I/II with Ki67 expression <20%. Basal-like tumors were negative for estrogen, progesterone, and HER2 receptors. We extracted the RNA and performed quantitative RT-PCR and statistical analyses (Table S3).

2.5 RNA extraction and quantitative RT-PCR from breast tumor samples

We used the miRNeasy mini kit (QIAGEN, Valencia, CA) for total RNA extraction from frozen tumors. We made cDNA by reverse transcription using Superscript II reverse transcriptase (Invitrogen, Carlsbad, CA) with 0.5–1 μg of RNA and 0.5 μg of random primers (Invitrogen) per reaction. For quantitative PCR analysis we used the 96-well plate Step One Plus system (Applied Biosystems, Carlsbad, CA) connected to an EpiMotion 5070 Robot (Eppendorf, Hauppauge, NY) and utilized SYBR Green PCR Master mix (Applied Biosystems). To verify primer efficiency we ran three serial dilutions of the indicated primer pairs (Table S3) in duplicate. We then normalized the quantity of mRNA from each gene to that of the human ribosomal protein large P0 (RPLP0) (de Cremoux et al., 2004). For each gene, we expressed the quantity “x” of that gene mRNA relative to the quantity of RPLP0 mRNA in a given sample by applying the formula x = 100*[Eˆ(Cp RPLP0 – Cp Gene)], where E is the mean efficiency of each primer pair. For statistical analysis we retained data from 69 patients for ASF1b, 70 patients for CAF-1 p60, 71 patients for CAF-1 p150, 67 patients for EZH2, 69 patients for MCM2 and 61 patients for HJURP, which fulfilled our amplification criteria (reproducible duplicates with consistent primer efficiency between samples).

2.6 Statistical analysis of breast tumors

We compared the differences in mRNA levels of each gene of interest with clinical factors including age, tumor stage, Elston-Ellis tumor grade, mitotic index, hormonal receptor status, HER2 status and triple negative carcinoma subtype, using an analysis of variance or the non parametric Wilcoxon's rank-sum test. We calculated correlations in mRNA expression levels between all genes of interest and between each gene of interest and the proliferation marker MKI67 by the Pearson correlation coefficient method. We compared gene expression levels in different breast cancer subtypes by using a two-sample Wilcoxon rank-sum test. We then defined a threshold value for each gene at the median mRNA expression level. Disease-free interval (DFI), disease-free survival (DFS), and metastasis-free interval (MFI) were estimated using the Kaplan–Meier method and each group was compared by log-rank and Wald tests. Briefly, DFI represents the length of time after tumor excision during which a patient survives with no relapse, defined by local recurrence, regional recurrence in lymph node-bearing areas, contralateral breast cancer and distant recurrences. DFS is similar to DFI but it considers the shorter period between the time to recurrence and the time to death for patients who experienced the two events, whereas MFI only takes into consideration tumor metastasis. A Cox proportional-hazards model analysis was performed to evaluate the prognostic value of each gene of interest in univariate analysis. We also carried out multivariate analysis to assess the relative influence of each prognostic factor tested on MFI, DFS, and DFI using a Cox stepwise forward procedure (Cox, 1972). For this analysis we used age, menopause, resection margin, mitotic index, HER2 and MKI67 (RNA and histological) receptor status, as well as CAF-1 p60, ASF1b, MCM2, and HJURP expression levels. We considered 59 patients in the multivariate analysis, as we considered only patients who did not have any missing value in any of the ten variables cited above. The significance level was set at 0.05 and statistical software R (2.11.0 version) was used for all our analyses. Corrections for multiple testing were done by the Benjamini–Hochberg test to control for false discovery rates.

2.7 Immunohistochemistry

We performed immunohistochemistry (IHC) on Adaptive Focused Acoustics (AFA) fixed, paraffin-embedded breast tumor sections mounted on precoated slides (Dako, Glostrup, Denmark), that were allowed to dry at 58 °C for 1hr. Sections were subjected to automated immunohistochemistry procedures using a LabVision autostainer (Dako). After deparaffinisation (in toluene) and rehydration (using ethanol at gradual dilutions) and antigen retrieval by immersion in a sodium citrate buffer (pH 6.0 for HJURP; pH 9.0 for MCM2 and Ki67) at 90 °C for 15 min, endogenous peroxidase activity was blocked by immersion in 3% H₂O₂ for 5 min, followed by a blocking step with 1% bovine serum albumin. Primary antibodies were incubated with the sections for 1 hr as indicated: polyclonal rabbit anti-HJURP antibody (1:100 dilution; ab100800; Abcam), monoclonal mouse anti-MCM2 antibody (1:100 dilution; MCA1859; AbD Serotec), and monoclonal mouse anti-Ki67 antibody (1:200 dilution; MIB-1; Dako). We used the EnVision+ Dual Link System-HRP (DAB+) polymer detection system with DAB+ as chromogen (Dako). Tissues were counterstained with hematoxylin for 2 min. Slide sections were acquired with a Philips ultra fast scanner and then images at ×40 and ×80 magnification were generated. We used a semi-quantitative intensity score (score 0: negative staining; score: 1, 2, 3: weak, intermediate, high nuclear staining and/or number of positive cells, respectively) for the analysis.

2.8 Ethics statement

Registration of patients of the Institut Curie in this cohort received a favorable agreement of the French National Committee on Computers and Liberties (CNIL, Commission Nationale de l'Informatique et des Libertés). According to French regulation, patients were informed of the types of experiments performed with the tumor specimens and did not express any opposition. The study was approved by the breast cancer study group and the clinical research study committee of the Institut Curie.

3 Results 3.1 Chromatin regulators cluster according to breast cancer molecular subtypes

We first aimed to assess the correlation between the mRNA expression of a broad selection of 54 chromatin regulators (CRs) (see materials and methods for selection details) and the four intrinsic breast cancer molecular subtypes. We also included ERBB2, ESR1, PGR, and MKI67 as controls. We performed unsupervised hierarchical clustering and principle component analysis of a published data set of 1127 patients (Reyal et al., 2008). We could cluster the four subtypes based on their similarities measured over the 58 genes (Figure 1B, Supp. Figure S1). As expected, ERBB2 (HER2) expression strongly correlated with the HER2+ subtype, and the estrogen and progesterone receptors showed the characteristic expression of hormone receptor positive subtypes (Figure 1B, Supp. Figure S1A, S1C) (Sorlie et al., 2001). Interestingly, we observed a gene cluster with high expression in the BLC, HER2+, and luminal B subtypes (red color) and low expression in the luminal A subtype (green color) (Figure 1C). This set comprised two histone variants, H2A.Z (Hua et al., 2008; Svotelis et al., 2010) and centromere protein A (CENP-A) (McGovern et al., 2012; Tomonaga et al., 2003), several histone chaperones, including two subunits of the chromatin assembly factor 1 (CAF-1) complex (p60 and p150), HJURP, and ASF1b, minichromosome maintenance (MCM) proteins (4, 6, 2), the Polycomb group protein enhancer of zeste homolog 2 (EZH2), and the proliferation marker MKI67. This data points to a CR putative “subtype gene signature” that clusters with luminal A tumors.

3.2 Chromatin regulators can distinguish breast cancer molecular subtypes

To investigate the ability of our chromatin regulators to significantly distinguish between the four molecular subtypes, we further selected seven CR genes from our putative “subtype gene signature” (HJURP, CENPA, ASF1b, CAF-1 p150, CAF-1 p60, MCM2, EZH2), and included MKI67 as a reference gene, and performed pairwise comparisons. Consistent with our clustering analysis, we found that the mean mRNA expression of all CRs and MKI67 was lower in the luminal A subtype compared to all other subtypes (Figure 2). We then assessed whether our CRs and MKI67 could differentiate between the BLC, HER2+, and Luminal B subtypes. Statistically, while MKI67 can nearly set apart BLC and HER2+ tumors (p = 0.05), all seven CRs tested in this cohort can significantly differentiate between at least two additional subtypes (Figure 2). Thus, our data shows CRs have the potential to clinically distinguish between the four intrinsic molecular subtypes.

Significant associations between the mRNA expression of select chromatin regulators and breast cancer molecular subtypes. Box plots represent the mRNA expression levels (logarithmic) from our microarray data set (n = 1127) in different breast cancer subtypes. Boxes represent the 25th–75th percentile; brackets: range; black line: median; black dots: outliers. mRNA expression comparisons among different subtypes were done first by ANOVA test and then by Wilcoxon rank-sum test. Significant p-values ([less than]0.05) were corrected for multiple testing by the Benjamini–Hochberg method and are indicated for each subtype pair. Basal-Like (BLC; n = 214), HER2+ (n = 144), luminal B (Lum B; n = 257), and luminal A (Lum A; n = 512).

To validate our findings, we exploited an independent patient cohort of 71 cryopreserved breast carcinoma samples (herein: ‘validation set’). This population is well annotated and features similar clinical characteristics (Table S2). Indeed, in contrast to the microarray data set, these tumors were node-negative and metastasis-free invasive breast carcinomas of a size that enabled breast-conserving surgery. We note that due to a limited amount of cDNA, we could not test all seven CRs in our gene signature. We noticed that a published report shows that the RNA expression of HJURP significantly correlates with the expression of its dedicated histone variant CENPA (Hu et al., 2010). Indeed, we found a significant correlation in the mRNA level of HJURP and CENPA in our microarray data set (Supp. Figure S2A) and transcriptome data sets from several cancers (Supp. Figure S2B). Thus, we chose to only retain HJURP for validation analyses. Importantly, similarly to the microarray cohort, the mean mRNA expression of nearly all factors was significantly lower in the luminal A subtype compared to the other subtypes (Figure 3). The two exceptions were ASF1b and CAF-1 p150, which could not distinguish HER2+ from luminal A tumors (Figure 3). Notably, while MKI67 and HJURP significantly differentiated luminal A from the other three subtypes, HJURP also significantly differentiated HER2+ from luminal B tumors (Figure 3). Together, our data show that HJURP, CAF-1 p60, MCM2, and EZH2 are markers to help classify breast cancer molecular subtypes.

Significant associations between the mRNA expression of select chromatin regulators and breast cancer molecular subtypes. Box plots represent the expression levels (logarithmic) from our validation set (n = 71) in different breast cancer subtypes. Boxes represent the 25th–75th percentile; brackets: range; black line: median; black dots: outliers. Comparisons of mRNA expression between subtypes were done by a two-sample Wilcoxon rank-sum test. Significant p-values from the test ([less than]0.05) were corrected for multiple testing by the Bonferroni method and are indicated for each subtype pair. BLC (n = 11), HER2+ (n = 6), Lum B (n = 16) and Lum A (n = 38).

Recent studies have challenged the reliability of Ki67 as a proliferation marker to differentiate borderline luminal A and luminal B patients, a critical step to decide therapeutic routes for ER+ patients. Even the most experienced laboratories utilize different cutoffs of Ki67 to classify patients (Piccart-Gebhart, 2011; Polley et al., 2013). We therefore asked whether using the median mRNA expression of our CRs could set apart luminal A and luminal B tumors, thus offering improved tools to differentiate ER+ patients. To address this question, we divided the microarray data set in two groups (high vs. low) according to their expression relative to the median value of each CR and determined their distribution in luminal A and luminal B subtypes. In line with our data above, luminal A patients had significantly lower mRNA levels of all factors compared to luminal B patients, determined by the Chi-square test (Table 1). Importantly, we observed the same trend in our validation set, with the one exception being CAF-1 p150 (Supp. Table S4). Collectively, our data highlight the benefit of using the median mRNA expression of our CRs as a standard cutoff to distinguish luminal A from luminal B tumors, and suggest that these represent new clinically valuable markers to improve breast cancer tumor substratification and better guide clinical decisions.

mRNA expression comparison (high vs low) of selected chromatin regulators in luminal B and luminal A subtypes (microarray data set).

3.3 HJURP, CENPA, and ASF1b are independent markers of disease progression

We next investigated the prognostic value of our seven CRs and compared it to that of MKI67. In our microarray data set, we divided the whole population into two groups using the median mRNA expression level of each factor as a cutoff. Univariate analysis revealed a significantly higher-risk of death (breast cancer specific survival, BCSS) associated with high mRNA levels of all seven CRs and MKI67 (Table S5). Similarly, high expression of HJURP, CAF-1 p150, MCM2, EZH2, CENPA and MKI67 was prognostic of metastasis (distant metastasis-free survival, DMFS) (Table S5). We then performed multivariate analysis in our microarray data set (Tables S6 and S7). We note that the high correlation of CENPA and HJURP can alter the significance of each factor if they are both included in the multivariate analysis, a phenomenon commonly observed in this type of statistical analysis (Tabachnick and Fidell, 1996). Thus, we performed the analysis with and without CENPA. Excluding clinical variables and CENPA, multivariate analysis revealed that the sole gene expression variable with independent prognostic value for BCSS was MKI67 (p = 0.001) (Table S6B), while only HJURP was prognostic of DMFS (p < 0.001) (Table S7B). When we included CENPA, we observed that CENPA had independent prognostic value for BCSS (p < 0.001) (Table S6C) and DMFS (p < 0.001) (Table S7C).

We next addressed the prognostic value of the selected factors and clinical variables using the validation set. At ten years, the overall survival, disease-free survival, disease-free interval, and metastasis-free interval were 81% [72–92%], 61.5% [50–75%], 67.5% [56.5–80.5%], and 84% [75–94%], respectively. Univariate analysis showed that mitotic index, tumor grade, HER2+ status (Tables 2–4), and high expression levels of HJURP, MCM2, and ASF1b were indicators of poor disease outcome when analyzing MFI, DFI, and DFS in the overall population (Supp. Figure S3). Notably, only histological Ki67 (p = 0.03) (Table 4), but not MKI67 mRNA expression (Supp. Figure S3), significantly associated with DFS. To test for independent prognostic factors, we performed a multivariate analysis adjusting for clinical and gene expression variables that were significant in the univariate analysis. In a forward stepwise analysis, menopausal status (p = 0.017), HER2+ status (p = 0.002), and high ASF1b mRNA levels (p = 0.023) were the only independent factors predictive of metastasis (Table 2). Moreover, only high HJURP mRNA levels were independently prognostic for DFI (p = 0.006) (Table 3) and DFS (p = 0.004) (Table 4). Taken together, the data from our validation set support the results from our microarray data set and show that in the overall population, the mRNA levels of ASF1b and HJURP are independent prognostic factors for disease outcome, whereas MKI67 RNA is not.

Uni- and multi-variate Cox regression analysis for metastasis-free interval, entire patient cohort (validation set).

Uni- and multi-variate Cox regression analysis for disease-free interval, entire patient cohort (validation set).

Uni- and multi-variate Cox regression analysis for disease-free survival, entire patient cohort (validation set).

3.4 HJURP is an independent marker of disease outcome in luminal A patients

Finally, we investigated whether HJURP, CENPA, ASF1b, MCM2, and MKI67 could predict disease outcome within the individual molecular subtypes. In the microarray data set, univariate analysis revealed that in the luminal A subtype, high mRNA levels of HJURP were prognostic of metastasis (i.e. DMFS) (p < 0.006; n = 316), while we were unable to find significant associations between the mRNA expression of any factor and BCSS in any subtype. Further, multivariate analysis considering only gene expression variables also indicated that in the luminal A subtype, high expression of HJURP was an independent prognostic factor for DMFS (HR = 2.49; 95%CI [1.29-4.84]; p = 0.007; n = 316). In our validation set, univariate analysis showed that high mRNA expression of HJURP and MCM2 significantly associated with a greater likelihood of developing metastasis in the luminal A subtype (p = 0.003 and 0.01, respectively; n = 38) (Figure 4 ‘top’). Moreover, high expression of HJURP and MCM2 significantly associated with DFI (p = 0.001 and 0.04, respectively; n = 38) (Figure 4 ‘middle’) and DFS (p = 0.005 and 0.02, respectively; n = 38) (Figure 4 ‘bottom’). Multivariate analysis revealed that HJURP was the sole independent prognostic factor for MFI (HR = 15.75; 95%CI [1.42-174.93]; p = 0.003), DFI (HR = 9.08; 95%CI [1.80-45.89]; p = 0.001), and DFS (HR = 6.56; 95%CI [1.44-29.78]; p = 0.005) in the luminal A subtype. Notably, the expression of MKI67 failed to significantly associate with MFI, DFI, or DFS in our analyses (Figure 4). Taken together, our microarray and validation datasets corroborate that luminal A patients with tumors expressing high mRNA levels of HJURP will have a shorter overall survival and a higher probability of developing metastasis. Thus, we demonstrate that the histone chaperone HJURP is, to the best of our knowledge, the first marker (based on gene expression analysis) that can distinguish good and poor prognosis patients within the luminal A subtype, representing an important tool to aid patient substratification and treatment choices for this breast cancer subtype.

High HJURP and/or MCM2 expression significantly differentiates between good and poor outcome luminal A patients. Univariate Kaplan–Meier curves of metastasis-free interval, disease-free interval, and disease-free survival in patients with low or high mRNA expression of the indicated gene using the validation set. A cut-off value that divided the population into two was set according to median mRNA expression, as indicated. The blue line represents high mRNA expression while the red line represents low mRNA expression. The number of patients at risk per time point is shown below each plot. Only significant associations between a gene and a molecular subtype are shown, the p-values (p [less than] 0.05) were obtained from a log-rank test and are in red. Expression of MKI67 is shown for comparison. BLC (n = 11), HER2+ (n = 6), Lum B (n = 16) and Lum A (n = 38).

4 Discussion

The vast heterogeneity of breast carcinomas and their different responses to treatment stress the importance of proper patient stratification to aid clinical decisions (Reis-Filho and Pusztai, 2011). The four major markers, estrogen receptor, progesterone receptor, HER2 status, and the proliferation marker Ki67, enable a first level classification. However, in certain cases these markers do not address the risk of recurrence or are sufficient to distinguish borderline populations (Patani et al., 2013). While histological Ki67 remains a valuable marker, new tools exploiting gene expression analyses that aid clinical decision makers are becoming more standardized and cost-effective (Cobleigh et al., 2005; Paik et al., 2004; van 't Veer et al., 2002; van de Vijver et al., 2002). Interestingly, new tests including Mammaprint and Oncotype DX, for example, largely ignore chromatin regulators. Indeed only CENPA and MCM6 are included in the 70-gene signature of Mammaprint while the Oncotype DX 16-gene signature does not consider a single CR (Paik et al., 2004; Tian et al., 2010). Furthermore, new studies challenge the clinical value of Ki67 at the RNA level as a standardized biomarker to distinguish luminal A and luminal B tumors and its prognostic value at the protein level compared to multi-level gene signatures (Polley et al., 2013; Tobin et al., 2014). Thus, we reexamined available transcriptome data to first identify a putative “gene signature” based solely on CRs that could distinguish breast cancer molecular subtypes and second, we report a factor, namely the histone chaperone HJURP, which can differentiate favorable and unfavorable outcome patients within luminal A breast carcinomas.

4.1 Selected CRs as a luminal A “subtype gene signature”

Clustering the expression of 54 CRs in a cohort of 1127 patients (Reyal et al., 2008) revealed a putative “subtype gene signature” with low expression in the luminal A subtype and high expression in the other three proliferative subtypes (Figure 1). Many of the factors in our putative “gene signature” have been implicated in breast and/or other cancers (Bachmann et al., 2006; Collett et al., 2006; Corpet et al., 2011; Gerdes, 1990; Giaginis et al., 2010; Hu et al., 2010; Hua et al., 2008; McGovern et al., 2012; Polo et al., 2004; Svotelis et al., 2010; Tomonaga et al., 2003). Indeed, ASF1b is overexpressed in many tumors (Abascal et al., 2013) and has prognostic value for metastasis in breast cancer (Corpet et al., 2011). The histone chaperone HJURP is overexpressed in lung, glioma, astrocytoma, and breast cancer (de Tayrac et al., 2013; Hu et al., 2010; Kato et al., 2007; Valente et al., 2013), while CENPA is overexpressed in many of the more aggressive tumors (Gu et al., 2014; Li et al., 2011b; Ma et al., 2003; Qiu et al., 2013; Tomonaga et al., 2003; Wu et al., 2012). Additionally, the methyltransferase EZH2 is overexpressed in certain cancers but it is also subject to gain-of-function or loss-of-function mutations that have made EZH2 a popular therapeutic target (Crea et al., 2012). Additionally, finding HJURP, CENPA, and the histone variant H2A.Z together in the cluster is revealing, given their importance for centromere maintenance (Boyarchuk et al., 2014, 2011), where misregulation of these proteins can cause chromosome segregation defects, genome instability, and possibly aneuploidy (Amato et al., 2009; Dunleavy et al., 2009; Foltz et al., 2009; Kato et al., 2007; Mishra et al., 2011).

4.2 Chromatin regulators are more than proliferation markers

Several meta-analyses using large sample cohorts suggest that the prognostic ability of most gene expression signatures relies primarily on detecting proliferation activity (Reyal et al., 2008; Wirapati et al., 2008). Our data showing a similar expression profile between selected CRs and MKI67 (Figure 1) might suggest that the CRs in our putative “gene signature” are solely additional proliferation markers, and thus offer information redundant to that of MKI67. While we also found significant correlations between select CRs and mitotic index (Figure S4; >10 mitosis) and MKI67 (Figure S5), our data highlights additional advantages of CRs. For example, in our microarray and validation data sets, MKI67 fails to significantly differentiate between BLC, HER2+, and Luminal B patients (Figures 2, 3), while all seven of our CRs can distinguish at least two other breast cancer subtypes in addition to luminal A in the microarray dataset (Figure 2). Furthermore, CRs were statistically better prognostic markers than both mitotic index and histological Ki67 in univariate and multivariate analyses (Tables 2–4), and HJURP outperformed MKI67 for prognostic value within the luminal A subtype (Figure 4). Indeed, this initial result has motivated us to pursue similar analyses with larger cohorts and use new analytical tools with higher resolution, such as Nanostring and SureSelect technology from Agilent. Taken together, our data suggest that our putative CR signature brings additional information to the clinic that could complement routinely used markers such as KI67, ER, PR, and HER2.

4.3 The value of CRs warrants consideration at the protein level

Here we focused on the mRNA expression of a selection of CRs. However, it is also important to consider that the expression at the RNA level does not necessarily imply similar expression at the protein level. In the clinical setting, immunohistochemistry is a commonly utilized test to measure proteins levels, including those of the classical marker Ki67. Interestingly, the use of histological Ki67 as a clinical tool remains controversial, particularly regarding its analytical validity and reproducibility across different laboratories (Polley et al., 2013), and in borderline cases with varying cutoffs (Varga et al., 2012). We therefore wanted to explore whether we could visualize our CRs by IHC in the different breast cancer molecular subtypes, and whether they had potential clinical value at the protein level. Using antibodies against HJURP and MCM2, our preliminary analyses reveal a variable labeling of the nucleus and/or cytoplasm in tumor samples (Supp. Figure S6, Table S8). Notably, the proteins appear more abundant in the proliferative subtypes compared to luminal A, in agreement with our gene expression findings (Figures 1–3). Though these results are preliminary, this work and others that have analyzed CRs by IHC (Castellini et al., 2014; de Tayrac et al., 2013; Rajput et al., 2011; Tsourlakis et al., 2013), open avenues for future work to (i) optimize the conditions and antibodies for several CRs in our subtype gene signature, and (ii) analyze their expression in IHC applications and their prognostic value in a parallel manner in different molecular subtypes. Standardized procedures will be essential for proper applications in clinical settings.

Notably, the significant correlation in this cohort at the mRNA level of the expression of HJURP and CENPA (Supp. Figure S2A) is consistent with a previous report (Hu et al., 2010) and several cancers in The Cancer Genome Atlas (TCGA) (Supp. Figure S2B). Whether this correlation translates to the protein level is important to consider given the potential biological consequences of an imbalance in the level of histone chaperones and histone variants and the possibility for crosstalk (Lacoste et al., 2014; Ray-Gallet et al., 2011). We recently reported using a model cell line stably overexpressing CENP-A that CENP-A can localize outside the centromere in the chromosome arms. We found that under these conditions, CENP-A could interact not only with its dedicated chaperone HJURP, but also with the H3.3 chaperone DAXX. This crosstalk leads to a DAXX-dependent aberrant mislocalization of CENP-A to the chromosome arms and gives the cells a survival advantage to overcome DNA damage induced by two anti-cancer treatments, camptothecin and ionizing radiation (Lacoste et al., 2014). This cellular mechanism highlights the importance for future work to investigate first, whether the correlation of HJURP and CENPA at the RNA level observed in many cancers translates to the protein level, and secondly, the biological implications of a potential imbalance, including cytogenetically using mitotic spreads to locate CENP-A. Thus, we propose that the two genes should be integrated into clinical tests to help guide patient substratification and treatment decisions.

4.4 HJURP as a new luminal A prognostic marker

Luminal A patients are still often treated with chemotherapy and many will only suffer unnecessary side effects with no clinical benefits (Cheang et al., 2009). In this respect, clinicopathological factors remain suboptimal for guiding clinical decisions regarding adjuvant therapy (Harbeck et al., 2014). Thus, new markers are needed to address these critical issues. Here, we find HJURP as the first biomarker to distinguish good and poor prognosis patients within the luminal A subtype. This significant result has potential clinical applications, particularly if we consider that patients with a low-risk of recurrence after surgery can be spared from adjuvant therapy (e.g., chemotherapy), as they are unlikely to clinically benefit but will still suffer harsh side effects and unnecessary costs. We note that our analysis featured a limited patient number. To explore whether HJURP can distinguish good and poor prognosis luminal A patients in a larger sample size, we exploited a metadatabase (www.kmplot.com) (Gyorffy et al., 2010). We find HJURP expression is significant in DMFS, RFS, and OS (Supp. Figure S7), further supporting its use in the clinical setting. Future work should focus on the integration of HJURP, CENPA, and other CRs from our putative “subtype gene signature” with currently utilized clinical markers to better classify and/or substratify breast cancer tumors, and serve as diagnostic or prognostic tools to help guide clinical decisions in breast cancer and perhaps in other types of cancer.

5 Conclusions

We identified a specific chromatin regulator gene cluster that allowed us to distinguish luminal A from luminal B tumors. We also identified HJURP as the first biomarker that can differentiate good and poor prognostic luminal A breast cancer patients. Our data supports the integration of selected chromatin regulators in the clinical setting to help guide treatment plans and improve overall breast cancer patient management.

Funding

This work was supported by la Ligue Nationale contre le Cancer (Equipe labellisée Ligue and a postdoctoral fellowship to R.M. and Z.G.L.), the European Commission Network of Excellence EpiGeneSys (HEALTH-F4-2010-257082), ERC Advanced Grant 2009-AdG_20090506 “Eccentric”, the European Commission large-scale integrating Project FP7_HEALTH-2010-259743 “MODHEP”, ANR “ChromaTin” ANR-10-BLAN-1326-03, ANR-11-LABX-0044_DEEP and ANR-10-IDEX-0001-02 PSL*, ANR “CHAPINHIB” ANR-12-BSV5-0022-02 and Aviesan-ITMO Cancer Project “Epigenomics of breast cancer”.

Author's contributions

Conceived and designed transcriptome analysis: RM, FB, FR, GA. Conceived and designed qRT-PCR analysis: RM, AC, L de K, GA. Performed qRT-PCR experiments: HR. Provided patient samples or data sets: FR, XSG. Performed statistical analysis: FB, AS, LW, BA. Analyzed data: RM, ZGL, FB. Performed and analyzed IHC: EM, DM, IV, RM. Wrote the paper: ZGL, RM, GA. All authors read and approved the final manuscript.

Competing interests

The authors declare that they have no competing interests.

Acknowledgments

We thank the Almouzni laboratory in particular Jean-Pierre Quivy, also Paul Cottu and Laura Attardi for discussions and critical comments on this manuscript.

Supplementary data

Supplementary data related to this article can be found at http://dx.doi.org/10.1016/j.molonc.2014.11.002.