Centromere protein-A, an essential centromere protein, is a prognostic marker for relapse in estrogen receptor-positive breast cancer

To determine the prognostic ability of CENP-A, we used two separate datasets of clinically node-negative patients who did not receive systemic therapy. The first is a cohort of 289 patients previously described by Wang, et al. [18] with gene expression data publicly available at the Gene Expression Omnibus (GEO) database [GEO:GSE2034]. The second is a series of 198 patients previously reported by the TRANSBIG consortium [19] with gene expression data available at the GEO database [GEO:GSE7390]. To evaluate the ability of CENP-A to predict response to tamoxifen, we used a dataset of 276 patients from the Institut Jules Bordet (JBI) [10, 20, 21]. Expression data are available at the GEO database [GEO:GSE2990]. Finally, to determine the value of CENP-A as a predictor of response to neoadjuvant chemotherapy, the MDA233 dataset of 233 patients treated with either weekly (80 mg/m² for 12 doses) or once every three weeks (225 mg/m² for four doses) paclitaxel followed by four cycles of 5-fluorouracil (500 mg/m²), doxorubicin (50 mg/m²), and cyclophosphamide (500 mg/m²; T/FAC) was used. Response to chemotherapy was categorized using the residual cancer burden (RCB) index based on residual disease at the time of surgery. Patients were categorized as having no (RCB = 0), minimal (RCB = 1), moderate (RCB = 2) or extensive (RCB = 3) residual disease as previously described [22]. Gene expression data from this cohort have been previously described [23, 24] and are available at the MD Anderson Bioinformatics website [25]. For all four datasets, ER status was determined using ligand binding assay [18, 21], EIA [18], or immunohistochemistry [18, 19, 24]. A summary of dataset properties is shown in Figure 1. Patient characteristics for all four datasets are shown in Table 1.

For all four datasets, mRNA expression levels were measured using the Affymetrix HG-U133A Gene Chip (Santa Clara, CA, USA) using standard procedures as previously described [18, 19, 21, 24]. mRNA levels were normalized using the MAS5 method to a target intensity of 600 followed by log2-transformation [26]. After normalization and transformation, a one-point increase in log2-transformed value corresponds to a doubling of mRNA expression. The following Affymetrix U133A probe set identification codes were used: CENP-A, 204962_s_at; CENP-B, 212437_at; HER2, 216836_s_at; Ki-67, 212022_s_at; and ESR1, 205225_at.

Correlations between CENP-A, CENP-B levels and ER status; CENP-A and RCB; and CENP-A and grade were assessed by analysis of variance (ANOVA) tests.

To analyze patient outcomes, Kaplan-Meier survival curves [27] for distant relapse-free survival (DRFS) were generated using tertile levels of expression of CENP-A, as there are no published levels of CENP-A expression to use as cutoffs. All cutoff levels were determined prior to data analysis. Survival curves were compared using the log-rank test. For all comparisons, statistical significance was set at P < 0.05, and all P-values were two-sided.

Univariate and multivariate Cox regression analyses were performed to identify correlations between variables and DRFS up to 5 years for the Wang, TRANSBIG and JBI datasets. Age, estrogen receptor alpha (ESR1) gene expression level, CENP-A gene expression level, Ki-67 gene expression level, and Her2/neu gene expression level were treated as continuous variables. The reference vs. comparison states for categorical variables were, T1 vs. T2 to T4 for tumor stage, 1 to 2 vs. 3 for tumor grade, and negative vs. positive for nodal status. Regression analyses were not performed on the Wang dataset due to the substantial amount of unknown data, or for the ER-negative JBI cohort due to the low number of patients in this group.

Univariate and multivariable logistic regression analyses were performed to identify correlations between variables and chemotherapy response (RCB, as described above) for the MDA233 data set. Predictive variables used were the same as in the DRFS multivariate regression analyses above.

The R statistical environment was used for all calculations [28].

We began by determining the level of CENP-A in ER-positive versus ER-negative tumors. In the Wang, TRANSBIG, MDA233, and JBI datasets, the levels of CENP-A were significantly higher in ER-negative compared to ER-positive tumors (Figure 2).

We then compared the level of CENP-B in ER-positive versus ER-negative tumors. In all datasets, CENP-B was not different between ER-positive and ER-negative tumors (Figure 2). CENP-B did not correlate with CENP-A when analyzed over all patients and in the subsets of ER-positive and ER-negative patients (data not shown).

Because CENP-A is necessary for centromere specification, higher levels of CENP-A may simply reflect higher levels of cell division. Therefore, we determined the correlation coefficient between CENP-A and Ki-67 in ER-positive and ER-negative tumors (Table 2). In all four datasets of patients with ER-positive and ER-negative disease, CENP-A was positively and significantly correlated with Ki-67. There was minimal or no correlation between CENP-B and Ki-67 (Table 2).

Similarly, CENP-A may also act as a surrogate marker of grade. In the three datasets containing grade information, CENP-A was positively and significantly correlated with grade for ER-positive cases (Table 3). CENP-A was also positively correlated with grade for ER-negative samples in the TRANSBIG and MDA233 datasets (Table 3). There were only ten patients with ER-negative disease in the JBI dataset, so the analysis was not performed on this subset.

The median follow-up was 7.2 years in the Wang dataset, 10.0 years in the TRANSBIG dataset, and 6.5 years in the JBI dataset. Among ER-positive patients, CENP-A levels were significantly higher in patients with a distant relapse within five years in the untreated Wang and TRANSBIG cohorts and in the tamoxifen-treated JBI cohort (Figure 3). For ER-negative patients in the Wang and TRANSBIG datasets, there was no difference in CENP-A level between patients with and without a distant relapse at five years (data not shown).

There are no previously established cutoffs for high or low CENP-A levels, and the distribution of CENP-A levels did not show a clear cutoff in any of the datasets (data not shown). Therefore, patients were grouped by tertile of CENP-A expression. Among patients with ER-positive disease in the untreated Wang and TRANSBIG datasets, higher levels of CENP-A were consistently correlated with decreased DRFS (Figure 4). In the Wang dataset, patients with the highest tertile of CENP-A expression had 5-year DRFS rates of 51% compared to 68% and 83% for patients in the middle and lowest tertiles of CENP-A expression (P = 0.0005, Figure 4A). A similar pattern was found for the TRANSBIG dataset (Figure 4B), with 5-year DRFS rates of 68%, 91%, and 100% (P = 0.0001). Among patients with ER-positive disease who were treated with tamoxifen in the JBI dataset, patients in the highest tertile of CENP-A expression had 5-year DRFS of 68% vs. 86% and 90% for patients in the middle and lowest tertile (P < 0.0008, Figure 4C).

For patients with ER-negative disease in the untreated datasets, there was no difference in 5-year DRFS based on tertile of CENP-A expression. In the Wang dataset, the 5-year DRFS for patients in the highest, middle, and lowest tertile of CENP-A level was 69%, 48%, and 58% respectively (P = 0.5). In the TRANSBIG dataset, the 5-year DRFS was 81%, 73%, and 48% for patients in the highest, middle, and lowest tertile (P = 0.065).

To determine if CENP-A provides independent prognostic or predictive information for DRFS, univariate and multivariate analyses were performed with age, T stage, tumor grade, and nodal status as categorical variables and with ESR1, CENP-A, Ki-67, and Her2/neu gene expression levels as continuous variables. Among patients with ER-positive disease in the untreated TRANSBIG dataset, univariate analysis revealed that grade, Ki-67, and CENP-A were significantly associated with an increased risk of relapse at 5 years (Table 4). To explore the interaction of grade, Ki-67, and CENP-A, multivariate analyses of the untreated TRANSBIG dataset were performed with and without CENP-A (Table 4). When CENP-A was not included in the multivariate analysis, Ki-67 was significantly associated with DRFS (hazard ratio (HR) 2.83; 95% confidence interval (CI), 1.48 to 5.42; P = 0.0017) and grade was borderline significantly correlated with DRFS (HR 2.65; 95% CI, 0.93 to 7.55; P = 0.069). When CENP-A was included in the multivariate analysis, neither Ki-67 nor grade correlated with DRFS (Table 4).

Unexpectedly, ESR1 had a HR of 2.00 (95% CI, 1.30 to 3.08; P = 0.0015) for relapse on multivariate analysis of ER-positive tumors in the TRANSBIG dataset (Table 4). On univariate analysis, ESR1 was not associated with an increased risk of relapse (HR = 1.32; 95% CI, 0.87 to 1.98, P = 0.19; Table 4). This discrepancy results suggests that the significance of ESR1 seen on multivariate analysis of the TRANSBIG dataset was the result of a high-dimension artifact.

For patients with ER-positive disease in the tamoxifen-treated JBI cohort, CENP-A was significantly associated with 5-year DRFS on univariate analysis (HR 1.97, 95% CI 1.38 to 2.80, P = 0.00017) and borderline significant on multivariate analysis (95% CI, 0.99 to 2.71, P = 0.054; Table 5). For patients with ER-negative breast cancer in the TRANSBIG dataset, CENP-A was not significantly associated with DRFS on univariate (HR 0.79; 95% CI, 0.51 to 1.23; P = 0.30) or multivariate analysis (Table 6). Similarly, for patients with ER-negative disease in the MDA233 dataset, CENP-A was not significant for distant relapse on univariate (HR 1.38; 95% CI, 0.88 to 2.17; P = 0.16) or multivariate analysis (Table 6).

To determine if CENP-A was associated with response to neoadjuvant chemotherapy, the MDA233 dataset was queried to determine CENP-A levels in patients with no or minimal residual disease after chemotherapy (RCB = 0/1) and in patients with moderate or extensive residual disease after chemotherapy (RCB = 2/3). In patients with ER-positive and ER-negative disease, CENP-A levels were higher in patients with no or minimal residual disease (Figure 5). CENP-B levels were not different between patients in the RCB 0/1 or RCB 2/3 groups for either ER-positive or ER-negative patients (data not shown).

Regression analyses were performed to determine if CENP-A level independently correlated with response to chemotherapy. On multivariate analysis of patients with ER-positive disease in MDA233, nodal status was marginally associated (P = 0.05) with RCB 0/1 vs. RCB 2/3 (Table 5). Similar analysis of patients with ER-negative disease in MDA233 identified age and T stage as significantly associated with chemotherapy response (Table 6).