Centrosome amplification induces high grade features and is prognostic of worse outcomes in breast cancer

The breast cancer tissue microarray (TMA) used in this analysis has been described previously [26, 27]. Briefly, samples were obtained from primary breast tumor blocks obtained at time of surgery for stage I-III breast cancer patients seen at the University of Wisconsin Carbone Cancer Center under protocol OS10111. The University of Wisconsin Health Sciences Institutional Review Board approved the TMA creation and approved use of the TMA and the de-identified coded data set (IRB approval 2010-0405). This protocol retrospectively collected de-identified data and archived tissue; the IRB waived patient consent. The TMA contains three 0.6 mm punch biopsies from each patient’s tumor, and 15 normal breast controls from mammoplasty are included in the array. All cases had at least 5 years of follow-up or recurrence or death within 5 years. Clinical information includes age at diagnosis, ethnicity, tumor size, lymph node involvement, stage, estrogen receptor (ER), progesterone receptor (PR), and HER2 status, type of surgery, adjuvant breast cancer treatments, and follow-up data, including any recurrence and death. Clinical data was obtained from the UW Hospital and Clinics Cancer Registry and manual chart review. ER, PR, and HER2 immunohistochemistry were also performed on the completed TMAs and interpreted by a breast pathologist. If ER/HER2 clinical data was not available, the clinical pathologic data from the original tumor sample was used for analysis. Patients with unknown or equivocal values were excluded from these analyses of subtype and CA. For subtype analysis, the following groups were used based on their clinical relevance [28, 29]: ER or PR positive and HER2-nonamplified; HER2-amplified; and triple negative.

Breast cancer TMAs were sectioned at 5 μm thickness, deparaffinized, and rehydrated. Antigen retrieval was performed in a pressure cooker at 250 °F with citrate buffer (pH 6) for 4 minutes. Blocking was done for 1 hour in 10 % fetal bovine serum (FBS) in PBS. Tissues were probed with anti-pericentrin (Abcam, ab4448, 1:200) and anti-polyglutamylated tubulin (Adipogen, GT335, 1:100) antibodies diluted in 1 % FBS and 0.1 % triton X in PBS overnight in a humidified chamber at 4 °C. Pericentrin and polyglutamylated tubulin are bona fide markers of centrosomes [30‐32]. The TMAs were then incubated with anti-rabbit Alexa 488 and anti-mouse IgG1 Alexa 647 secondary antibodies (Jackson ImmunoResearch Laboratories, West Grove, PA) for 1 hour at room temperature. Slides were washed 3 times after primary and secondary antibody incubations. Slides were counterstained for DNA with 4′,6-diamidino-2-phenylindole (DAPI) and mounted with ProLong Gold antifade reagent (Life Technologies). Scoring of centrosome phenotypes was performed using a Nikon Eclipse Ti inverted microscope, 100x objective, and CoolSNAP HQ2 charge-coupled device camera (Photometrics). The observer was blinded to clinical data and analyzed centrosomes in a minimum of 30 cells per case from 3 different tumor regions. The number of distinct pericentrin foci as well as foci that overlapped with polyglutamylated tubulin were counted. Cell boundaries were visualized by nonspecific background staining with the polyglutamylated tubulin antibody. Average centrosome number per cell was calculated for each case. Centrosome sizes were measured in at least 15 representative centrosomes per case from three different tumor regions using the pericentrin marker, and an average was calculated for each case. For survival analysis, the median centrosome size (0.99 μm) was used as the cutoff for large versus small centrosomes. In addition to number and size, we also noted any unusual centrosome phenotypes such as centrosome clustering, centrosome speckling, and atypical shapes.

A small fraction of samples in the TMA were not evaluable due to loss of tissue, insufficient cellularity, or other technical issues and were excluded from analysis. Centrosome data were linked to de-identified clinical data by sample number and position on the TMA and sorted for analysis using Microsoft Excel.

Fluorescence in situ hybridization (FISH) was performed using standard techniques, as reported elsewhere [33]. Briefly, chromosomes 4, 10, and 17 were probed on one section, and chromosomes 3, 7, and 9 on another section. Chromosomes were counted by observers blinded to patient conditions in a minimum of 10 cells per case. A small fraction of samples were not evaluable due to loss of tissue, insufficient cellularity, or other technical issues and were excluded from analysis. Similarly a subset of samples had a single probe that was not well visualized, but if at least five chromosomes were available, it was included in further analyses. FISH data were linked to de-identified clinical data by sample number and position on the TMA and sorted for analysis using Microsoft Excel. Ploidy was determined by the average chromosome number for all 6 probes combined. CIN was determined as the average percentage of cells that deviated from the modal number for each of the 6 chromosomes assessed by FISH. Samples were considered to have CIN if this value exceeded 45 %, a cutoff that yielded appropriate percentages of normal samples and tumors with CIN.

The doxycycline-inducible PLK4^WT and PLK4⁶⁰⁸ MCF10A and RPE cell lines were a kind gift from Dr. David Pellman. Cells were cultured and centrosome amplification was induced as previously described [24, 33]. For assays, cells were treated with 2 μg/mL doxycycline for 48 hours and subsequently harvested for qRT-PCR and flow cytometry. Immunofluorescence was performed as previously described using the following antibodies: anti-pericentrin (Abcam, ab4448), anti-gamma tubulin (Abcam, ab27074), anti-alpha tubulin (Millipore, MAB1864), and Alexa fluorophore-conjugated secondary antibodies (Jackson).

RNA was isolated from cells using the RNeasy Micro Kit (Qiagen, Valencia, CA), and converted to cDNA using the Quantitect Reverse Transcription Kit (Qiagen). qRT-PCR was performed using EvaGreen master mix (MidSci, St. Louis, MO) and a StepOne Plus instrument (Applied Biosystems). Quantification of cytokeratins 7, 18, and 19 (KRT7, KRT8, KRT19) expressed as mRNA level was normalized to the mRNA of three housekeeping genes (RRN18S, GAPDH and ACTB). Primers sequences are provided in Additional file 1: Table S3. Fold changes in gene expression were assessed using the 2^-ΔΔCt method [34].

A total of 50,000 events were acquired for each sample using an Accuri C6 flow cytometer (Accuri, Ann Arbor, MI) equipped with multicolor analysis, and data were analyzed with Flow Jo 7.0 (Tree Star, Ashland, OR). Samples were run in triplicate in at least 3 independent experiments. The following antibodies were used: CD24-PE, CD44-PE, and mouse IgG1 isotype control (BD Biosciences). Mean channel fluorescence of FL2 was used to quantitatively compare conditions.

R (version 3.1.1, R Core Team, Vienna, Austria) statistical software was used for survival analysis. A total of 362 patients were included in survival analyses. The clinical outcomes analyzed in this study were recurrence-free survival (RFS) and overall survival (OS). RFS was defined as the time from initial breast cancer diagnosis to recurrence. OS was defined as the time from diagnosis to the date of death. RFS and OS were plotted using the Kaplan–Meier method, and log-rank tests were used to compare patients with tumors with CA versus tumors with no CA using 2 centrosomes per cell as a cutoff. Sensitivity analyses were also performed using the mean and median of all the normal breast samples in the TMA as the cutoff for defining CA. Cox proportional hazards model included centrosome amplification, stage, tumor grade, hormone receptor status, and HER2 status. Associations between these factors and either RFS or OS were analyzed and presented as hazard ratios (HR) with 95 % confidence intervals (CI). For centrosome size, an average size was calculated for each case. The median of all cases was used as the cutoff for the large versus small centrosome groups. The correlations of CA with ploidy and CIN were assessed with Spearman’s correlation. The correlations between centrosome amplification and grade or stage were performed by stratifying patients by grade or stage and comparing the mean centrosome number among the groups by Kruskal-Wallis tests. Non-parametric tests were used because the distribution of average centrosome number was right skewed (Additional file 1: Figure S1). Two-sided, unpaired statistical tests were used throughout. P < 0.05 was considered statistically significant for all statistical tests.

We initially characterized the distribution of centrosome abnormalities and clinical characteristics seen in our breast cancer samples. Patient characteristics are shown in Additional file 1: Table S1. Centrosomes were assessed in each sample using pericentrin, a PCM marker. The normal mammary gland is composed of terminal ductal lobular units, and polarity is well defined, as indicated by luminal positioning of the centrosome and basal positioning of the nucleus (Fig. 1a, top). However, this organization is disrupted in carcinoma samples (Fig. 1a, bottom). Furthermore, the median centrosome number from all the tumor samples was almost double that of the normal breast samples (1.8 vs. 1.0, p = 0.001), and 84 % (305 of 362) of breast cancer samples had a mean centrosome value higher than that of the normal breast samples. Additionally, the average centrosome number per cell and the percent of cells with greater than 2 centrosomes were both significantly greater in breast cancers compared to normal breast (Fig. 1b-c). To demonstrate that this is not simply due to a greater proliferative rate in the tumors (as centrosomes are duplicated at the G1/S transition), we correlated average centrosome number with Ki67, a marker of proliferative index. Although there is a partial correlation, a significant portion of samples have high centrosome number with low Ki67 (Fig. 1d). Because of the partial correlation, an elevated average centrosome number between 1 and 2 could indicate an increase in the percent of tumor cells in G2. Hence, we used a strict cutoff of >2 centrosomes for subsequent analyses of CA. Breast cancers also showed a wider distribution of mean centrosome number per cell compared to normal breast (range 0.5–5.9 and 0.23–2.77, respectively), consistent with CA occurring in a variable fraction of cells within a tumor. The distribution of centrosomes in breast cancer was unimodal with right skew, while the distribution of centrosomes in normal breast samples was normal (Additional file 1: Figure S1).

We stratified patients based on stage, grade, subtype, regional node status, and recurrence site. Patients with higher stage and grade also had a higher average number of centrosomes per cell (Fig. 2a-b). Furthermore, CA was greater in triple negative and HER2 amplified subtypes (Fig. 2c); in general, estrogen/progesterone receptor-positive breast cancers have a more favorable prognosis than HER2 amplified or triple negative cancers [35].

A tumor was considered to have CA if the mean number of pericentrin foci per cell exceeded 2. Using this definition, CA was found in 35.1 % of breast tumors and 13.3 % of normal breast samples. It was most common in triple negative breast cancers (61.4 %) and less frequent in HER2-positive (41.2 %) and hormone-sensitive/HER2-negative subgroups (29.2 %). We next assessed how CA correlated with clinical outcomes. Patients with CA had significantly worse overall survival (OS, P = 0.002; Fig. 2d) and recurrence-free survival (RFS, P < 0.001; Fig. 2e) than those without CA. Furthermore, these patients also had worse breast cancer-specific mortality (P = 0.003; Fig. 2f). Our findings led us to hypothesize that high-CA tumors may provide useful prognostic data in addition to providing a biologic reason for aggressive breast cancers. To be clinically useful, CA would need to indicate risk that is not captured with currently available clinical factors such as tumor stage, grade, and subtype. To test this, we performed Cox proportional hazards modeling (Additional file 1: Table S2). This analysis demonstrated that stage and hormone receptor status were the strongest predictors of OS and RFS. When corrected for these, CA is not an independent predictor of OS or RFS. Although CA does not provide a clinical factor independent of known risk factors, it nevertheless may provide a biological explanation for how tumors advance in grade and stage.

To survey additional centrosome defects, we observed aberrations in centrosome shape, size, and patterning (Additional file 1: Figure S2). Centrosome clustering was observed in 58 % of tumors versus 13 % of normal samples. Centrosome speckling (clusters with >5 centrosomes) was observed in 23 % of tumors versus 7 % of normal samples, and irregular centrosome shapes in 41 % of tumors versus 20 % of normal samples (this relatively high incidence of abnormal shapes in normal samples likely represents staining artifact). What we term centrosome speckling has been described by others as sand-like centrosomes [12]. We did not observe worse clinical outcomes with atypically shaped centrosomes or centrosome speckling, although centrosome clustering correlated with significantly worse OS (P = 0.009) and RFS (P = 0.030). Centrosome clustering has been proposed as a mechanism by which cells with CA are able to divide with pseudo-bipolar spindles [36, 37], although it is unclear whether the interphase clustering observed here would correspond with clustering during mitosis.

One potential mechanism leading to CA is cell-doubling events (e.g. cytokinesis failure, cell-cell fusion). If cell doubling represented the primary cause of CA in breast cancer, we would expect a strong correlation between CA and increased cell ploidy. Therefore, we evaluated how CA correlates with high tumor ploidy, as determined by 6-chromosome FISH in 354 breast tumors. Ploidy ranged from 1.43 to 8.75 with a median of 2.08. We find that CA strongly correlates with ploidy (P = 0.006; Fig. 3a). Further, after dividing patients by CA (defined as >2 centrosomes per cell), tumors with CA had significantly greater ploidy (Fig. 3b). To verify these findings, analyses were repeated using a more stringent definition for centrosomes: the overlap of pericentrin and polyglutamylated tubulin, which represents the overlap of PCM and centriole markers, respectively [30, 31]. Using these criteria, CA still correlated with ploidy (Additional file 1: Figure S3). These data provide evidence that CA and whole genomic amplifications occur in concert in incipient tumor cells, suggesting that genome doubling events occur commonly in breast cancer oncogenesis.

To estimate what percentage of CA events arise from doubling events, we calculated the percent of tumors with CA (average centrosome number >2) that also had elevated ploidy (>3). This revealed that at least 15 % of CA events arose from doubling events (Fig. 3c). However, this method is likely to underestimate the true percentage of CA events that arise from doubling events because cells that originate after genome doubling can subsequently lose chromosomes [38, 39].

CA can lead to multipolar cell division or lagging chromosomes through induction of merotelic attachments on focused bipolar spindles, resulting in chromosomal instability (CIN) [39]. Therefore, we examined the relationship between CA and CIN. CIN was calculated as the percent of cells within a tumor with a non-modal number of chromosomes, averaged for 6 chromosomes. 44.7 % of breast tumors have CIN compared to 9.1 % of normal breast samples. Patients whose tumors displayed CIN had worse breast cancer-related overall survival (Additional file 1: Figure S3). CA correlated positively with CIN (Fig. 3d, P < 0.001). After dividing the patients into two groups based the presence of CA, as done for survival analyses, tumors with higher CA had significantly elevated CIN (Fig. 3e). These data support the hypothesis that CA is a common cause of CIN in breast cancer. In addition, we found a strong positive correlation between ploidy and CIN (Fig. 3f).

As done above for CA and ploidy analysis, we repeated other analyses using the more stringent definition of pericentrin and polyglutamylated tubulin overlap. Similar to our analysis based on pericentrin staining alone, CA defined by the overlap of pericentrin and polyglutamylated tubulin was more pronounced in triple negative breast cancer and cancers with higher histological grade (Additional file 1: Figure S4A-D). Patients with CA had worse overall and recurrence-free survival (Additional file 1: Figure S4E-G). Furthermore, CA as defined by these criteria also demonstrated a significant correlation with ploidy and CIN (Additional file 1: Figure S5A-D). In summary, we observed similar findings whether centrosomes were defined using solely pericentrin or using the overlap of pericentrin and polyglutamylated tubulin.

Although centrioles are surrounded by PCM in normal cells, ~1/3 of cells in normal samples had PCM without detectable centrioles, suggesting that only a subset of centrioles were labeled with the polyglutamylated tubulin antibody. However, compared with normal samples, tumors more frequently had pericentrin foci that lacked co-staining with polyglutamylated tubulin. An average of 78 % of pericentrin foci contained this centriolar marker in normal samples compared to an average of 46 % in breast tumors. 302 out of 362 breast cancer cases had a percentage lower than 78 %, suggesting a true loss of this centriole marker in some breast tumors. These findings suggest that either these tumor centrioles lack polyglutamylated tubulin, or that acentriolar centrosomes are a bona fide characteristic of many human breast cancers. Acentriolar centrosomes have been reported previously in cancer cells and are thought to result from PCM fragmentation [19, 40]. Additionally, acentriolar centrosomes were more common in the triple-negative breast cancer subtype and correlated with advanced stage and grade (Fig. 4a-c), although there was no significant correlation with worse clinical outcomes (Fig. 4d-e; P = 0.202 for overall survival and P = 0.133 for recurrence-free survival). Nevertheless, these findings indicate that PCM fragmentation is potentially a marker of more aggressive tumors.

Because there was a strong correlation of CA with poorly differentiated tumors (grade 3) in our study and others [6‐8], we hypothesized that CA induces cellular de-differentiation. To test this, we utilized doxycycline-inducible PLK4 in MCF10A and RPE cell lines [24, 41], in which the overexpression of PLK4 results in CA (Fig. 5a, d), and subsequently looked at markers of differentiation. Breast cancer cells that express less CD24 and more CD44 are more de-differentiated and more stem cell-like [42‐44]. These cells may also have enhanced metastatic potential [45]. We analyzed CD24 and CD44 by flow cytometry after inducing CA in MCF10A cells, and found that this significantly decreased CD24 and increased CD44 (Fig. 5b-c). To ensure this was not an effect of doxycycline or of PLK4 expression independent of centrosome amplification, we employed a doxycycline-inducible PLK4^1-608 cell line in which this kinase is expressed without amplifying centrosomes due to lack of a critical localization domain. We did not observe markers of de-differentiation with doxycycline in this control (Fig. 5c). To validate this finding, we employed a second cell line, immortalized retinal pigment epithelial (RPE) cells, for which cytokeratin profiles can reveal differentiation status. More de-differentiated cells express excess cytokeratins 7 and 19 and less cytokeratin 18 [46‐49]. We assessed expression of these 3 cytokeratins by qRT-PCR, finding that RPE cells with CA express more cytokeratins 7 and 19, but less cytokeratin 18 (Fig. 5e), which is consistent with a de-differentiated state.

The Nottingham grading scale uses the following three criteria to determine the differentiation status of a tumor: (1) the amount of gland formation, (2) nuclear pleomorphisms, and (3) mitotic figures [50, 51]. To address whether CA is sufficient to impart these characteristics, we first observed which tumors demonstrated glandular/tubular structures in at least two of three histologic regions examined. Indeed, tumors from the TMA without tubule formation had greater CA (Fig. 5f). With regard to the second criterion, previous work has shown cells with CA can exhibit multipolar spindles and other nuclear pleomorphisms, and this is seen in vitro as well (Fig. 5g). For the third criterion, it has already been demonstrated that cells with CA proliferate more slowly [24, 52], which would tend to cause lower grade tumors; however, numerical CA positively correlates with Ki67 status in our data set (Pearson r = 0.3106, p < 0.001, Fig. 1d) suggesting that CA does not cause tumor cells to exit the cell cycle. Furthermore, it has been demonstrated that cells with CA take longer to complete mitosis due to multipolar spindle formation [53], which could explain why more mitotic figures are seen in tumors with CA. Taken together, these data support the idea that CA directly or indirectly imparts high-grade features to tumors, leading to worse clinical outcomes.

All of the data on which the conclusions of the paper rely are included in the main figures or supplementary figures.