A Quantitative Centrosomal Amplification Score Predicts Local Recurrence of Ductal Carcinoma In Situ

Approximately 20% of screen-detected breast cancers are ductal carcinoma in situ (DCIS), a preinvasive form of breast cancer wherein malignant epithelial cells are confined to the lumen of a mammary duct and do not invade into the adjacent stroma (1, 2). Notably, 20% to 53% of women with untreated DCIS progress to invasive breast cancer (IBC) over a period of ≥10 years (3). Because the progressive potential of a DCIS lesion cannot be reliably determined, local control via surgical excision with or without local radiotherapy (RT) is the mainstay strategy, with addition of endocrine blockade in some cases (4). Unfortunately, 10% to 35% of patients with DCIS treated with lumpectomy or breast conservation surgery (BCS) later present with a local recurrence (LR), and about half of all recurrences occur in the form of IBC (5, 6). A major challenge is to avoid under- or overtreatment by developing prognostic biomarkers that can stratify patients with DCIS based on their recurrence risk.

Current predictors of recurrence risk for DCIS such as the Van Nuys Prognostic Index (VNPI; ref. 7) and the Memorial Sloan Kettering DCIS nomogram (8) are based on routinely used clinicopathologic parameters but lack consistency and reproducibility in risk prediction (9, 10). In addition, these tools do not integrate prognostically informative molecular predictors and underestimate DCIS heterogeneity. Although Oncotype Dx Breast DCIS score, a commercially available gene expression–based assay, has some value in predicting LR, it has only been validated in two cohorts (ECOG E5194 and Ontario DCIS). The poor stratification of high/intermediate-risk patients in these two cohorts has called into question the prognostic value of this tool (11).

Extensive genetic and phenotypic intratumoral heterogeneity (ITH) characterizes DCIS (12, 13). In a preinvasive lesion, higher ITH predicts greater likelihood of LR and IBC (14). Amplified centrosomes underlie erroneous mitoses and fuel chromosomal instability (CIN), which is a well-recognized driver of ITH (15, 16). Although normal cells have one centrosome pre–S-phase and two centrosomes post–S-phase, cancer cells invariably display centrosome amplification (CA), an abnormal increase in the number (i.e., numerical amplification) and/or volume (i.e., structural amplification) of centrosomes (17). Semiquantitative studies have shown that CA correlates with higher tumor grade, larger tumor size, disease recurrence, and/or distant metastasis in various malignancies (18). Moreover, CA occurs within precancerous and preinvasive lesions including DCIS, suggesting that CA is an early event in tumorigenesis (19, 20). CA increases with higher DCIS grade, and high-grade DCIS has elevated expression of Aurora-A and Nek2 kinases that are strongly associated with CA. In addition, the risk of LR in DCIS is predictable by dysregulation of genes like cyclin-D, cyclin-E, and p53/p21 that regulate the centrosome duplication process (21). In the present study, we postulated that recurrent and nonrecurrent DCIS cases might differ in the extent and/or type of CA. The prognostic value of CA has remained unexplored for clinical application, as there is no methodology available for the rigorous quantitation of CA phenotypes. Also, it is unclear whether the prognostic value of CA lies in numerical and/or structural CA. It is unknown which of the two features of CA—frequency (i.e., percentage of cells showing amplified centrosomes) and/or severity (i.e., how abnormal the number/volume of centrosomes is in a given sample)—is prognostically informative.

Herein, we present a new methodology for centrosomal phenotyping to quantitate both numerical and structural centrosomal aberrations in clinical tissue samples. Centrosomes were immunofluorescently stained using an antibody against γ-tubulin and costained nuclei with Hoechst. Our analytical procedure allows robust interrogation of the capacity of centrosomal overload to predict the risk of LR after a lumpectomy. We have developed an algorithm that quantitates the frequency/prevalence and severity of CA (both numerical and structural) in formalin-fixed paraffin-embedded (FFPE) clinical samples, and computes a centrosome amplification score (CAS) for each sample. CAS is a promising metric that may improve treatment recommendations and allow identification of patients at low risk of recurrence for whom adjuvant RT may not be required. CAS demonstrates the highest concordance among the known prognostic models such as VNPI and commonly used clinicopathologic variables such as grade, age, and comedo necrosis.

FFPE tissue sections of primary pure DCIS consecutively diagnosed between 1988 and 2012 were obtained for this retrospective study from Nottingham City Hospital, UK. Tumor tissues were preserved by standard approved processing methods using formalin fixation and embedding in paraffin. These tumor blocks were stored in the Nottingham tissue bank. Patients who had (i) adequate amount of tissue, (ii) available all relevant clinicopathologic data, and (iii) at least 10 years of follow-up were eventually included in the study. The samples for the study were shared in three batches. For the pilot study to estimate the sample size, samples for the first 50 consecutive cases that met inclusion criteria were shared, and based upon our findings, the proposed sample size of 116 for each cohort was expected to yield a power of 80% with an alpha of 0.05 (Supplementary Fig. S1). Subsequently, samples for the next 83 cases were shared which together with the earlier 50 samples formed the discovery cohort (DC). The validation cohort (VC) was received only after the study (staining, imaging, and image analysis) on the DC was completed. To exclude any bias, the Georgia State University (GSU) research group was totally blinded to clinicopathologic and outcome details of the patients included in the study. These data were not shared with the GSU research team who performed the staining, imaging, and image analysis until the CAS scores were generated for each patient in all cohorts. The DC (n = 133) and VC (n = 119) comprised of consecutive pure DCIS patients (no evidence of microinvasive or IBC) with available tissue samples that showed free surgical margins >2 mm (to avoid the effect of this confounder on the study outcome) and underwent BCS or mastectomy with or without adjuvant RT (Supplementary Fig. S2; refs. 22–24). All cases were histologically reviewed, and diagnoses were confirmed by two independent pathologists (M.S. Toss and I.M. Miligy), and in case of disagreement between the two reviewing pathologists, a specialist breast pathologist (E.A. Rakha) confirmed the diagnosis. All cases included data pertaining to their clinicopathologic variables such as age at diagnosis, menopausal status, DCIS size, nuclear grade, presence of comedo-type necrosis, treatment, VNPI, Ki67 proliferation index, and information about treatment (adjuvant RT), recurrence-free survival (RFS) defined by the time (in months) between 6 months after the first surgery and occurrence of ipsilateral LR in the form of either DCIS or IBC, date of initial diagnosis, date of surgery, and patient status at last contact (23). Patients who underwent completion surgery within the first 6 months after primary resection surgery due to positive/close surgical margins or presence of residual tumor tissue were not considered to have disease recurrence. All patients who developed contralateral breast events were censored at the time of development of the contralateral tumor. None of the patients in our DC and VC received adjuvant endocrine therapy.

To determine normal volumes of the centrosomes, full-face sections of normal breast tissue from reduction mammoplasties (n = 40) and breast tumors with extensive regions of adjacent uninvolved tissues (n = 40) were obtained from Stavanger University Hospital, Norway; Nottingham City Hospital, UK; and West Georgia Hospital, GA. All aspects of study were (i) approved by every Institutional Review Board, (ii) in compliance with guidelines on material transfer and data use agreements for all involved institutions, and (iii) conducted in accordance with International Ethical Guidelines for Biomedical Research Involving Human Subjects. Informed consent was obtained from all subjects.

Centrosomes were immunofluorescently stained for γ-tubulin (red) and nuclei (with Hoechst) in paraffin-embedded sections of DCIS. Images of tissue samples were acquired with a Zeiss LSM 700 confocal microscope (using 63× oil immersion lens with a numerical aperture of 1.4 at 1.5× optical zoom). All imaging parameters were fixed across all samples. For optimal results, laser power was adjusted to the minimum level wherein fluorophore emission was saturated. For detector saturation, the gain (master) was adjusted such that the detector registers the target fluorophores in each channel within full range of detector settings (8-bit, 12-bit, 16-bit) to prevent over- and undersaturation and maximize accuracy. The offset was adjusted to minimize the background in the sample. Normal, DCIS, and IBC areas premarked by a pathologist were imaged to obtain at least 10 regions of interest (ROI), with each containing 20 to 30 nuclei and associated centrosomes (Fig. 1).

Raw three-dimensional (3D) image data were processed using IMARIS Biplane 8.2 3D volume rendering software to determine the volume of each centrosome within each ROI. “Volume rendering” refers to transforming a two-dimensional (2D) image stack for 3D visualization and subsequent analysis. To exclude nonspecific signals, a common background subtraction was applied to all images. This parameter was derived by first measuring the average diameter of approximately 100 centrosomes in 10 ROIs (Fig. 1), and then using the background corresponding to this average diameter as the background subtraction threshold. Finally, data from all optical sections were ordered to enable volume measurement for each centrosome. The final data of volumes of all centrosomes were then compared with a maximum intensity projection image, and centrosomes for each cell were quantified based on proximity to their associated nuclei. The number and volume of all centrosomes associated with each nucleus in the tumor area were recorded.

Centrosomes in breast tissue (normal, DCIS, or IBC) were categorized into individually distinguishable centrosomes (iCTR) and megacentrosomes (mCTR). iCTRs were defined as centrosomes that stain positive for γ-tubulin; iCTR numbers and boundaries were clearly distinguishable, and their volumes lay within the range of centrosome volumes found in normal breast tissue stained for γ-tubulin. The volume range for a normal centrosome was determined by analyzing volumes of centrosomes from both adjacent uninvolved tissue from patients with cancer and normal breast tissue from disease-free individuals (Supplementary Fig. S3). For adjacent uninvolved tissues, the selected cohort (n = 40 patients) had a median age of 53.5 years (age range, 38–69.5 years). We evaluated centrosomal volumes in these samples as described in the analysis section. The mean centrosome volume for the adjacent uninvolved tissue sections was higher relative to the normal tissue from reduction mammoplasty. Thus, we chose the smallest and largest values for individual centrosome volume from normal tissue as the “normal centrosome volume range” for breast tissue. The mean volume of centrosomes in normal breast epithelial cells ranged from 0.2 to 0.74 μm³. Centrosomes with volumes > 0.74 μm³ were categorized as mCTRs. All centrosomes in each ROI were thus categorized as iCTRs or mCTRs. In other words, mCTRs are centrosomes with aberrantly large volumes and are considered to represent structurally amplified centrosomes. The numbers and volumes of each iCTR and mCTR associated with each nucleus in an ROI were recorded.

For each sample, a cumulative CAS (CAStotal) was computed based on the formula: CAStotal = CASi + CASm, where CASi and CASm are scores that describe numerical and structural CA phenotypes, respectively. Details on quantitation of numerical and structural CA are added in Supplementary Data.

Statistical analysis was accomplished with SAS 9.4 software and the R-project version 3.4.3 (R Foundation for Statistical Computing, https://www.R-project.org/). Raw CA volume data were converted to CASi, CASm, and CAStotal according to the algorithm. Scaling factors recommended were used to normalize score of CASi and CASm in the range 0 to 3. The χ² tests were performed to check recurrence proportions in patient subgroups. The tests of group mean differences shown in Box plots were based on nonparametric Wilcoxon Rank Sum Tests and Kruskal–Wallis tests depending on the number of groups used for comparison, where the y axis reflects the ranks of observations. RFS was used as the endpoint for the survival analysis (restricted to 10 years). The optimal cutoff (threshold used to categorize patients into high or low risk of LR subgroups) of the CAStotal value was selected based on the results of 133 log-rank tests. We simply set each possible CAStotal value from 133 cases in the DC as cutoff and then constructed Kaplan–Meier survival estimators for cases classified into high-risk and low-risk groups. The value 1.436 was finalized because it minimized the log-rank P value. The same CAStotal cutoff was then used for the 119 cases from the VC to validate the model's effectiveness. Both univariate and multivariable Cox proportional hazard models, with age, grade, comedo necrosis, and RT controlled, were built to estimate HRs and 95% confidence intervals between high versus low CAStotal groups. A nonzero slope was detected in a generalized linear regression of the scaled Schoenfeld residuals on functions of time, which satisfied the proportional hazards assumption (Supplementary Fig. S5). A 2 × 2 confusion matrix and performance metrics was used for sensitivity analysis. The fitted Cox models were also used to predict the approximate 10-year recurrence rate (RR) using SAS PROC PHREG module. For all tests, P < 0.05 was considered to be statistically significant.

We found that among the 133 patients in the DC (details in Table 1), 28 patients developed ipsilateral LR. The median age at diagnosis was 58 years (range, 41–84 years), and median follow-up was 132 months (range, 14–333 months). Out of 133 patients, approximately 42% (n = 55) received RT. Higher nuclear grade, the presence of comedo necrosis, and the use of RT were clinicopathologic parameters that showed proportional differences between recurring and LR-free patient subgroups (Table 1A). However, only high-grade and comedo necrosis showed associations with RFS in a univariable Cox regression analysis (Table 2A). Intriguingly, none of these clinicopathologic variables showed any significant association with RFS in multivariate analyses (Table 2A), thereby indicating the limited capacity of traditional clinicopathologic variables to predict LR for DCIS in our DC. Our VC was also from Nottingham University Hospital, UK (patient characteristics in Table 1B), and comprised of 119 patients with DCIS out of which 24 patients presented with ipsilateral LR. Median age of these patients was 56 years, and the median follow-up was 121 months. Histograms representing distribution of age and tumor size are added in the Supplementary Data (Supplementary Fig. S6). In addition, we performed the Kaplan–Meier survival analysis to show the effect of standard prognostic markers like age, tumor size, RT, and comedo necrosis on recurrence for the whole dataset (DC and VC, n = 252; Supplementary Fig. S7). Out of 119 patients, approximately 12% (n = 14) received RT. In the VC, tumor size, presence of comedo necrosis, and age showed significant proportional differences between the LR and LR-free subgroups.

Centrosome numbers and volumes, evaluated and scored for numerical (CASi) and structural (CASm) centrosomal aberrations (as described in Materials and Methods), were integrated using our algorithm to generate a composite CAStotal value for each sample of the DC (Fig. 2A and B). Interestingly, patients with DCIS that developed LR within 10 years showed significantly higher CASi relative to LR-free patients (P = < 0.0001; Fig. 2C). These patients with LR showed greater severity (CASi severity; P = 0.25; Supplementary Fig. S8A) and higher frequency (CASi frequency; P < 0.0001; Supplementary Fig. S8B) of numerical CA compared with LR-free patients. Analysis of structural CA revealed that CASm was significantly higher (P = 0.04, Fig. 2D) for the LR subgroup relative to the LR-free subgroup. DCIS with LR exhibited greater severity (CASm severity; P = 0.01, Supplementary Fig. S8C) and frequency (CASm frequency; P = 0.08, Supplementary Fig. S8D) of structural CA compared with LR-free DCIS. Cumulatively, a summation of CASi and CASm generated CAStotal, which was significantly higher for patients with DCIS with LR relative to LR-free patients regardless of grade (mean scores in Supplementary Table S1; Fig. 2E).

Employing the same methodology for the VC, we calculated CAS (Supplementary Fig. S9) and found that irrespective of grade, DCIS cases with LR exhibited higher CAStotal relative to LR-free patients (P < 0.0001; Fig. 2F). Further, similar trends were seen for other CAS subcomponents as observed in the DC; the ranked mean values of CASi (P < 0.0001; Fig. 2G) and CASm (P < 0.0001; Fig. 2H), including their severity (CASi severity P = 0.0014; CASm severity P = 0.014) and frequency (CASi frequency P < 0.0001, CASm frequency P < 0.0001) components, were higher in the patient subgroup with LR than in the LR-free subgroup (Supplementary Fig. S8E–S8H).

Similar findings were evident for grade-matched patients in DC and VC (Supplementary Fig. S10) and patients that were treated only with BCS (Supplementary Fig. S11). Collectively, our data strongly suggest a stark difference in centrosomal aberrations between DCIS tumors of patients with and without LR.

Next, we coimmunolabeled 15 high-grade DCIS samples for both centrosomes (using anti–γ-tubulin antibody) and centrioles (using anti–centrin-2 antibody) and generated CAStotal as described before. In all samples, γ-tubulin foci invariably overlapped with centrin-2 foci, confirming that both structurally and numerically amplified centrosomes are bona fide centrosomes and not simply aggregates of pericentriolar material. We also observed that none of the mCTRs had >2 centrin-2 foci, suggesting that enlarged γ-tubulin foci represent structurally augmented centrosomes and not supernumerary centrosomes that are tightly clustered to be indistinguishable (Supplementary Fig. S12).

Upon stratification of all DC patients into low- and high-CAS groups (the threshold used was the one that minimized log-rank P value; Fig. 3), we found that patients with DCIS with high CASi were associated with poorer RFS (P < 0.001, HR = 4.80) relative to those with low CASi (Fig. 3A; Supplementary Fig. S13A and S13B; Supplementary Table S2). Similarly, high CASm was associated with poorer RFS (P = 0.04, HR = 2.396) compared with low CASm (Fig. 3B; Supplementary Fig. S13C and S13D; Supplementary Table S2). CAStotal stratified the high-risk and low-risk DCIS patients with high significance and HR (P < 0.001, HR = 6.3; Fig. 3C). We found that 85.7% of patients with LR were in the high CAStotal group. This association with CAStotal remained significant (P < 0.001, HR = 7.4) even after accounting for potential confounders, including comedo necrosis, tumor grade, age, RT, and receptor status (Table 2A). Although presence of comedo necrosis and CAStotal was associated with RFS in univariate analyses, only CAStotal remained significantly associated with RFS in multivariable analyses (Table 2A). Furthermore, when similar Cox regression univariate and multivariate analysis was performed for CASi and CASm separately, CASi and CASm was the strongest and most significant independent predictor of RFS, respectively (Supplementary Tables S3A and S4A). Similar results were evident for the cases that were treated only with lumpectomy (Supplementary Fig. S14).

To verify whether CAStotal, CASi, and CASm could be used to stratify patients in the VC, we used predetermined CAS cutoffs from the DC (Fig. 3). We found that high CASi, CASm, and CAStotal were associated with poorer RFS compared with low CASi, CASm, and CAStotal, respectively. Of the patients with LR, 75% were classified into the high-CASi group (Fig. 3D) and approximately 67% of patients with LR were classified into the high-CAStotal subgroups (Fig. 3E). Of the patients in the recurrence-free group, 87% were classified in the low-CASm group (Fig. 3F). In both univariate and multivariate analyses after adjusting for potentially confounding effects of factors like age, grade, RT, and receptor status, CAStotal and comedo necrosis was the strongest and most significant independent predictor of RFS (i.e., HRs for CAStotal were higher than HRs of all other clinicopathologic factors; Table 2B). Similar to DC, we observed that CASi and CASm also independently predicted the RFS (Supplementary Tables S3B and S4B).

In addition, we performed the bootstrap analysis for the Cox regression univariate and multivariate models on the combined (DC+VC = 252) dataset and observed that mean HR for the univariate analysis is 5.22 and the multivariate analysis conditional on all other variables is 6.58 (P < 0.0001; Supplementary Fig. S15; Supplementary Table S5). Also, CAStotal was able to identify patients for both DCIS (Supplementary Fig. S16A and S16B; Supplementary Table S8Ai and S8Bii) and invasive recurrence even after adjusting for potentially confounding effects of factors like age, grade, and RT (Supplementary Fig. S16C and S16D; Supplementary Table S8AiI and S8Bii) in both DC and VC (clinicopathologic characteristics summarized in Supplementary Tables S6 and S7).

Further, in both the DC and VC, the 10-year estimated risk of LR increased continuously as the CAS increased (Supplementary Fig. S17). Next, we determined if our survival model had high predictive accuracy using the Harrell's concordance index. The higher the concordance index, the better the survival model discriminates between patients who experienced LR versus those who remained LR free. The results indicated that any patient with a poorer/shorter RFS had a 72.6% probability of being in the high-CAStotal group. Also, we created 2 × 2 confusion matrix performance metrics to show the accuracy of CAS to predict 10-year LR. To do so, we calculated the sensitivity (Sn), specificity (Sp), positive predictive value (PPV), negative predictive value (NPV), and accuracy (Acc) of CAS and odds ratio (OR, which represents the increase in odds of a patient in a high-risk group developing recurrence relative to a patient in a low-risk group), for both cohorts to compare the performance of CAS with that of the traditional clinicopathologic variables (those used in the Cox regression analysis). As presented in the tables below, our CAStotal yielded an Acc of 0.60, Sn of 0.85, Sp of 0.53, PPV of 0.32, NPV of 0.93, and OR of 6.8 in the DC (Supplementary Table S9). We noticed that the CAStotal produced a lower accuracy and specificity compared with comedo necrosis (0.71). However, comparison of the Sp, PPV, NPV, and OR performance metrics showed the overall superiority of CAStotal, in both cohorts, when compared with the clinicopathologic variables.

Thus, these results collectively show that CAS can robustly predict 10-year LR risk for patients with DCIS from two different cohorts.

In the DC, CAStotal stratified patients with DCIS treated with surgery (mastectomy/BCS) or BCS alone (Supplementary Fig. S18B and S18C) into subgroups with high- and low-LR risks with greater significance relative to patients treated with surgery (mastectomy/BCS) and postoperative RT (Supplementary Fig. S18A; HR = 11.6, P < 0.0001 for surgery alone; HR = 17.05, P = 0.0005 for BCS alone; and HR = 2.4, P = 0.3589 for surgery + RT). Similarly, in the VC, CAS stratified patients with DCIS treated with surgery only (Supplementary Fig. S19A and S19B) into subgroups with high- and low-LR risks with higher significance compared with patients treated with surgery (mastectomy or BCS) and postoperative adjuvant RT (surgery+RT; HR = 3.97, P = 0.049 for surgery alone and HR = 1.4, P = 0.109 for surgery+RT). These data suggest that CAStotal can identify LR patients who might benefit from adjuvant RT. In addition, we observed that patients with DCIS who recurred as IBC exhibited higher CAStotal (P = 0.07) compared with the patients who recurred as DCIS (Supplementary Fig. S20) in the DC.

We next evaluated the clinical significance of CAS by examining the associations of CAS with traditionally employed clinicopathologic variables, i.e., age, grade, tumor size, comedo necrosis, and RT (Supplementary Figs. S21 and S22). Our data show that CAStotal provides clinically relevant prognostic information over and beyond what is provided by current clinicopathologic parameters alone. Given that high CA is associated with more aggressive disease phenotypes, we not only observed the association of high CAStotal with higher RRs, but also found that CAStotal segments patient subgroups more deeply than traditional clinicopathologic parameters (see RR forest plot in Supplementary Fig. S16A). For example, the RR forest plot (Supplementary Fig. S23A) for patients with high-grade DCIS in the DC showed that patients with comedo necrosis (red) are at high risk of recurrence (0.59) compared with the overall RR for patients (0.33), regardless of the CAS of their tumors. When we further stratified these patients with DCIS with comedo necrosis into high (green) and low (blue) CAS groups, we observed that the RR for the high-CAS group (green) was 0.83 and RR for the low-CAS group (blue) was 0.10. Similar results were observed for VC (see RR forest plot in Supplementary Fig. S23B). Thus, CAS was able to more deeply segment the patients with comedo necrosis into high- and low-risk LR groups. Similar trends were evident for tumor size, RT, and age.

The widely used VNPI is based on patients' age at diagnosis, tumor size, resection margin width, and tumor grade. To test the performance of this index in our (DC and VC combined) cohort, we calculated VNPI based on scoring methods described in the literature. Each of the factors was assigned a score between 1 and 3, and the sum of scores for the four parameters (i.e., the final VNPI score) was used to stratify patients into high-, low-, and intermediate-risk groups for LR, employing the binary cutoff score of ≥8. Next, we compared the performance of VNPI and CAStotal in cases from the DC and VC (n = 164; Fig. 4A and B) using univariate and Kaplan–Meier survival analyses. We found that higher VNPI was not significantly associated with poorer RFS, and VNPI did not significantly stratify patients as high and low risk of LR subgroups. By contrast, CAStotal stratified DC and VC patients into subgroups of high and low risk of LR with greater significance and HRs (CAStotal HR = 5.6 vs. VNPI HR = 0.70; Supplementary Table S10). Multivariable analyses adjusted for other potentially confounding factors, such as tumor size, presence of comedo necrosis, age, and RT along with VNPI and CAS, revealed that CAStotal showed the highest association with RFS, with an HR = 6.86 (Supplementary Table S11). These findings compellingly suggest that the CAS stratification of patients with DCIS is superior to that of the traditional VNPI index.

DCIS exhibits considerable interpatient heterogeneity and has a poorly understood natural history. A lack of accurate models for prediction of risk of LR results in over- and undertreatment, complicated by the variable prognostic evidence of patient age, tumor margins, DCIS grade, and size. CA is a hallmark of cancers and is observable in > 80% of breast tumors including preinvasive lesions, and is associated with high grade in DCIS and IBC (18, 19). Amplified centrosomes are present in premalignant cells and increase as the disease progresses to dysplasia, highlighting the potential involvement of CA in neoplastic transformation and progression (25).

Our laboratory has previously shown that (i) high levels of CA are associated with poor progression-free survival in invasive breast tumors, and (ii) CA is higher in the aggressive triple-negative breast cancer (TNBC) subtype compared with grade-matched non-TNBCs (26, 27). This notion was further validated by analysis of the CA20 gene score, which is based on genes associated with CA (28). Recent studies have reported that higher CA induces high-grade features in breast cancers; thus, CA has been associated with tumor evolution (29). Although studies have reported that breast cancers exhibit structurally amplified centrosomes, they have not yet established the prognostic value of this structural CA (30). This may be due, in part, to the 2D (i.e., cross-sectional) approaches used in these studies, which have limitations to accurately capture the 3D size of the centrosome. More so, most studies (31) examining CA in breast cancers have not rigorously evaluated confounding effects of other clinicopathologic variables on the prognostic value of CA.

Our new semiautomated methodology uses quantitative centrosomal phenotyping and an algorithm to measure both numerical and structural centrosomal aberrations in DCIS tumors. For each sample, a continuous CAS was computed that categorized patients as having a high or low 10-year risk of LR. Findings from our retrospective study, which involved two large, well-characterized cohorts (DC and VC) of DCIS cases, showed that patients with LR within 10 years exhibited higher CAStotal relative to LR-free patients. Our study is the first to show that organellar-level differences distinguish patients with DCIS with LR from LR-free patients, and that high levels of both numerical and structural CA are associated with increased 10-year risk of LR in patients with DCIS. Our results suggest that aberrant centrosomal homeostasis in DCIS drives pathophysiologic alterations that potentially facilitate disease progression through CIN-dependent as well as CIN-independent mechanisms. Although CA may drive ITH through CIN, an increased centrosome complement may, via modulation of the microtubule cytoskeleton, enhance directional migration and invasion of malignant cells and thus enhance the risk of LR in the longer term (32). We have demonstrated that CAStotal is significantly and independently associated with poor RFS, and upon inclusion of both CAStotal and VNPI into multivariable models, we found that CAStotal outperforms VNPI in predicting LR. CAStotal predicts the 10-year risk of LR with higher concordance than VNPI. In subsets of patients with DCIS, defined based on their clinical and histopathologic parameters, stratification by CAStotal prognostically augmented several clinicopathologic parameters in determining rate of recurrence. Among subsets of patients with DCIS treated with BCS or those receiving additional adjuvant RT, CAStotal identified patients with high risk of LR. Thus, CAStotal can be used as a clinical tool to identify patients who can be safely treated with BCS/mastectomy alone, and those who will benefit from the inclusion of RT. Our centrosomal profiling methodology, which dichotomizes patients with DCIS into high- and low-risk categories, enables clear go/no-go therapeutic decision-making and can substantially augment individualized management of DCIS based upon risk conferred by the patient's centrosomal complement.

CAS, as the linear expression of the severity and frequency of numerical and structural CA, may serve as an indirect measure of ITH in DCIS. Our study, the first to robustly quantify CA in both pure and mixed DCIS samples, has contributed evidence supporting a model of CA-driven DCIS progression into IBC. These findings concur with previous studies wherein we, and others, observed that TNBC, the most aggressive subtype of breast cancer, exhibits highest CA among all breast cancer subtypes (26, 29). Centrosome profiling can complement clinicopathologic and genomic evaluation to provide a comprehensive portrait of disease status. An exciting avenue for future research is to profile CA in all the stages of tumor progression starting from atypical hyperplasia to invasive and metastatic disease to evaluate if CA can function as a biomarker for tumor evolution.

The commercially available Oncotype Dx DCIS score is applicable mainly to cases with resection margins of at least 3 mm and low-/intermediate-grade DCIS measuring ≤2.5 cm, or in high-grade DCIS of ≤1 cm, as this is the set of patients from the ECOG 5194 study upon which the test was initially clinically validated (11). By contrast, our quantitative centrosomal phenotyping methodology is more broadly applicable and could be refined for other cancer types with rampant CA. The gene signature that comprises the basis of the Oncotype DCIS Score consists mainly of proliferation-related genes. CA is a phenotypic biomarker that serves as a readout of hundreds of deregulated signaling pathways that culminate in numerical and/or structural CA, including dysregulated proliferation-related signaling cascades. Thus, our methodology captures prognostic information from a broader swath of biological pathways that are deregulated in and drive the biology of DCIS. CAS-based risk profiling of core biopsies may reduce the number of re-excisions even in the event of close/positive margins.

However, our study has a few limitations. There are imbalances in the number of patients in different subgroups, in the DC and the VC of the study, which has resulted in better performance of CAS (higher HR) in the DC. Although the DC has more high-grade patients, the VC has a balanced number of high-, intermediate-, and low-grade patients. High-grade patients tend to present with invasive recurrence. A higher number of patients recurred as invasive in the DC and patients with invasive recurrence exhibited higher CAS when compared with patients who recurred as DCIS in DC. Whereas, in VC due to more balanced numbers of high-, intermediate-, and low-grade patients, no such variation in the type of LR was observed. Furthermore, lack of receptor status in some cases precluded study of the confounding effect of receptors in this dataset. The study cohort did not include any patients treated with endocrine therapy. These limitations in the DC and VC perhaps lead to the slightly different performance of CAS among the two cohorts. Validation studies in external cohorts and mechanistic studies to understand the role of CA-associated proteins in DCIS progression model are warranted.

K. Gogineni is a paid advisory board member for Sanofi Aventis and Pfizer. P. Rida is an employee for and holds ownership interest (including patents) in Novazoi Theranostics and is a paid consultant for Neo-Biz Solutions. No potential conflicts of interest were disclosed by the other authors.

Conception and design: K. Mittal, R.M. Osan, U. Manne, P. Rida, R. Aneja

Development of methodology: K. Mittal, G. Wei, R.M. Osan, P. Rida, E.A. Rakha

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): K. Mittal, M.S. Toss, B.D. Melton, I.M. Miligy, A.R. Green, E.A.M. Janssen, H. Søiland, E.A. Rakha

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): K. Mittal, G. Wei, D.H. Choi, R.M. Osan, E.A. Rakha

Writing, review, and/or revision of the manuscript: K. Mittal, M.S. Toss, G. Wei, R.M. Osan, I.M. Miligy, A.R. Green, E.A.M. Janssen, H. Søiland, K. Gogineni, U. Manne, P. Rida, E.A. Rakha

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): K. Mittal, M.S. Toss, D.H. Choi, I.M. Miligy, E.A.M. Janssen, H. Søiland, U. Manne

Study supervision: K. Mittal, R. Aneja

Other (fluorescent immunostaining, imaging): J. Kaur

This study was supported by grants to R. Aneja from the National Cancer Institutes of Health (U01 CA179671) and a graduate fellowship to K. Mittal from the Second Century Initiative Program at Georgia State University.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.