Cytoplasmic Cyclin E Predicts Recurrence in Patients with Breast Cancer

Overexpression of low molecular weight cyclin E (LMW-E), which was originally discovered in breast cancer by our group and subsequently reported by others (1–7), is also found in ovarian cancer (8, 9), melanoma (10), colorectal cancer (11–14), lung cancer (15), and renal cell carcinoma (16). LMW-E is generated from N-terminal neutrophil elastase cleavage of the 50-kDa, full-length cyclin E1 (17). The loss of the NH2 terminus alters the subcellular localization of LMW-E to the cytoplasm, which can no longer be degraded by the nuclear FBW-7, resulting in stable and functionally active LMW-E in the cytoplasm (18). LMW-E is not found in normal tissues, and it causes cell-cycle deregulation and resistance to endocrine therapy and chemotherapy (19–22). Compared with full-length cyclin E, LMW-E has a higher affinity for binding cyclin-dependent kinase 2 (CDK2) and is capable of phosphorylating substrates, such as pRb, with higher activity (23). Transgenic mice expressing LMW-E have a higher incidence of mammary tumor development and metastasis compared with those expressing full-length cyclin E (24), and tumorigenicity is dependent on CDK2 (25).

We previously reported an analysis of 395 patients with breast cancer in which we demonstrated that overexpression of LMW-E, as measured by Western blot analysis, is associated with distant metastases and reduced overall survival (26). At a median follow-up of 4 years, 91.7% of patients with overexpression of cyclin E (full-length and LMW-E) according to Western blot analysis developed distant metastases compared with 7% of patients without cyclin E overexpression (P < 0.001). In multivariable analysis, total cyclin E expression was an independent prognostic variable and a better predictor of both disease-free and overall survival than nodal status, estrogen receptor (ER) status, or stage.

Western blot analysis is impractical for routine clinical use, which has led several groups to examine the prognostic value of cyclin E expression measured by IHC. In these studies, cyclin E expression was scored on the basis of nuclear staining, and no consistent association between cyclin E overexpression and reduced survival was demonstrated (27). We previously reported that LMW-E is localized predominantly in the cytoplasm, where it binds to CDK2 and has greater kinase activity than full-length cyclin E (18). Others have reported that mislocalization of other cyclins, including cyclins A, B, and D1, has functional implications in tumorigenesis (28–31). These data suggest that examining tumors for expression of cyclins may be best assessed by IHC to determine location (nuclear vs. cytoplasmic) and intensity of staining. To test this hypothesis, we recently used separate nuclear and cytoplasmic scoring systems for both cyclin E and p-CDK2 expression to demonstrate altered cellular accumulation of these proteins using immunohistochemical analysis (32). We evaluated 7 different cyclin E antibodies to identify the one antibody that can consistently detect cytoplasmic cyclin E in a panel of 14 breast cell lines and tumors from cyclin E transgenic mouse models and showed that nuclear versus cytoplasmic staining of cyclin E readily differentiated full-length from LMW-E, respectively. Finally, we showed that cytoplasmic cyclin E correlated strongly with cytoplasmic p-CDK2 (P < 0.0001) in patient samples (32).

In this study, we tested the hypothesis that expression of LMW-E, as measured by cytoplasmic staining of cyclin E in IHC, predicts recurrence and survival outcomes in patients with breast cancer across diverse large cohorts and subtypes.

Four cohorts were analyzed in this study (Supplementary Table S1): a prospective cohort of 303 patients with stage I–III breast cancer who enrolled in our study between January 2000 and June 2010 (Lab00-222; approved by The University of Texas MD Anderson Cancer Center Institutional Review Board) and 3 retrospective cohorts, including 725 patients with stage I–II breast cancer treated at MD Anderson (Houston, TX) between 1985 and 1999 (MDA; ref.33), 951 patients with stage I–II breast cancer treated at four U.S. hospitals between 1985 and 1997 (NCI) and 515 elderly patients with stage I–II breast cancer treated at Nottingham University Hospitals, Nottingham, England, between 1987 and 2005 (UK). A complete database of all the clinical and pathologic factors evaluated for all 4 cohorts is provided in Supplementary Table S2. Fresh tumor and normal breast tissue was collected during surgery from 152 patients in the Lab00-222 cohort, and formalin-fixed, paraffin-embedded (FFPE) tissue was available for all patients in all four cohorts. REMARK diagram for each cohort is shown in Supplementary Fig. S3.

Fresh tissues from the 152 patients in the prospective cohort were collected and protein lysates subjected to Western blot analysis using an mAb to the C-terminus of cyclin E (HE12; Santa Cruz Biotechnology) and actin (Roche) as described previously (26, 34) with the following modifications. Briefly, 50 μg of protein from each tumor tissue being examined was loaded into the lanes of a 10% SDS-PAGE gel and submitted to electrophoresis for 300 V hours. These gels were then transferred to Immobilon P membrane (Millipore) by electrophoresis for 400 V hours at 4°C. The membranes were blocked for 1 hour in BLOTTO (5% nonfat dry milk in TBS-T, 20 mmol/L Tris, 137 mmol/L NaCl, 0.25% Tween, pH 7.6). Following blocking, the membranes were incubated in primary antibodies (at a concentration of 0.1 μg/mL) in BLOTTO for 3 hours at room temperature. Following incubation with primary antibody, the membranes were washed three times for 20 minutes in TBS-T and then incubated with goat anti-mouse or anti-rabbit IgG–horseradish peroxidase-conjugated secondary antibodies (Pierce) at a dilution of 1:3,000 in BLOTTO for 1 hour. The membrane was washed three times for 20 minutes in TBS-T and developed with the Renaissance chemiluminescence system (PerkinElmer Life Sciences, Inc.). The membranes were placed in an autoradiography cassette, exposed to film, and scanned. Densitometry was performed using ImageQuant Total Lab software (Amersham Biosciences), each band quantified, and LMW-E bands were added together. Values of LMW-E and full-length cyclin E were normalized against full-length cyclin E in normal tissue.

Slides from all patients within each cohort were stained for cyclin E using C-terminal antibody detecting full-length cyclin E and LMW-E (rabbit polyclonal antibody to cyclin E, C-19, sc-198; Santa Cruz Biotechnology) as recently described (32). Briefly, sections (5 μm thick) of FFPE breast tumor samples were deparaffinized and rehydrated followed by incubation with 3% hydrogen peroxide and methanol to block endogenous peroxidase activity and nonspecific protein–protein interactions, respectively. Antigen retrieval was carried out with 0.01 mmol/L citric acid–based buffer at pH 6.0 using a hot plate in a metal container for 15 minutes before immunostaining. After 1-hour blocking for nonspecific staining, the sections were incubated with cyclin E C-19 polyclonal antibody. Antibodies were detected using a VECTASTAIN Elite ABC Kit (PK6101 and PK6102; Vector Laboratories). Cells positive for cyclin E were visualized using the chromogenic substrate 3,3′-diaminobenzidine. Slides were counterstained with hematoxylin and mounted. Tumor cell blocks known to highly express LMW-E were included in each batch, and negative control was prepared by replacing the primary antibody with PBS buffer. All washing steps were performed with, first, PBS alone and then PBS with 0.1% Tween.

Cyclin E scoring was performed by three pathologists (CK, AB, OA) blinded to patient outcomes. Scores (0 = negative, 1 = weak staining, 2 = moderate staining, and 3 = strong staining) were assigned for nuclear and cytoplasmic staining according to percentage of cells stained and intensity of staining (Supplementary Fig. S1) and as described previously (32). Each tumor sample was scored separately for nuclear and cytoplasmic cyclin E expression, and LMW-E status was assigned as follows: LMW-E negative (no staining or just nuclear staining), LMW-E positive (nuclear + cytoplasmic or just cytoplasmic staining).

ER status, progesterone receptor (PR) status, HER2 status (3+ in IHC or amplified in FISH), and Ki67 status (35) were determined from pathology reports for the MD Anderson and UK cohorts. For the NCI cohort, we stained tissue samples for ER, PR, and HER2; FISH was performed at Quintiles. Samples were collected prior to the change in American Society of Clinical Oncology/College of American Pathologists guidelines for ER reporting (36); therefore, we considered >10% staining on IHC to be positive for ER and PR.

ER, PR, and HER2 were used to approximate breast cancer subtypes. ER⁺ and/or PR⁺ and HER2⁻ samples were considered hormone receptor–positive breast cancer; HER2-positive (independent of ER and PR status) was considered HER2-positive breast cancer; and ER⁻, PR⁻, and HER2⁻ samples were considered triple-negative breast cancer (Supplementary Table S1).

Logistic regression was performed to determine whether full-length cyclin E and LMW-E expression as detected by Western blot analysis were associated with nuclear and cytoplasmic staining for cyclin E in IHC. Differences in the cohorts were evaluated using χ² tests for categorical variables and ANOVA for Ki67. Endpoints were overall survival and breast cancer freedom from recurrence (FFR) calculated from the time of diagnosis to recurrence. FFR is a modification from the recurrence-free survival endpoint from the Hudis and colleagues' guidelines (37) and was calculated as the time between the date of diagnosis and the date of first recurrence (locoregional or distant) or last follow-up (if no recurrence). FFR captures only recurrences and does not include deaths as events, regardless of cause of death. Patients who did not experience the endpoint were censored at last follow-up.

Median follow-up times were computed using the reverse Kaplan–Meier estimator (38). The Kaplan–Meier method was used to calculate 5-year FFR for each factor. The differences in survival curves were evaluated using log-rank test. Univariable and multivariable regression analyses were performed using Cox proportional hazards model. The final multivariable Cox models were selected using backward elimination procedure by Akaike information criterion (AIC), which takes into account how well the model fits the data as well as the complexity of a model, thereby reducing the risk of overfitting. Model performance was quantified using concordance index (C-index; ref.39), which is a measure of the probability of agreement between what the model predicts and the actual observed risk of breast cancer recurrence. The variable importance in a model is evaluated using the likelihood ratio test (LRT). Statistical analyses were performed using R version 3.2.3 with library survival and rms. R markdown files described for each cohort are presented in Supplementary R Files S1–S4. Each R code includes the definitions of FFR and covariates used for each cohort, including subtype and cyclin E. The multivariable Cox proportional hazards models for each cohort were initially fit with the following covariates: age, T stage, nodal status, tumor grade, subtype, chemotherapy, radiotherapy, endocrine therapy, and cyclin E. Next, backward elimination procedure was performed for variable selection. These files also include calculations of c-indices and provide codes to draw monograms.

There were 2,494 patients with complete data available: 303 from Lab00-222, 725 from MDA, 951 from NCI, and 515 from UK. Clinical and pathologic variables for individual cohorts and the combined cohort are summarized in Table 1. The cohorts were distinct (P < 0.001) with respect to all variables examined except for cytoplasmic cyclin E staining. Patients in the UK cohort were older than in other cohorts (median 76 years compared with 53–62 years; P < 0.0001). In addition, T stage, nodal status, tumor grade, and subtypes were significantly different among the cohorts (P < 0.0001). Treatment variables (chemotherapy, radiotherapy, and endocrine therapy) were also significantly distinct among cohorts (P < 0.0001). In the UK cohort, none of the patients received chemotherapy. These results show that although the four cohorts were distinct in all biomarkers examined, they converged with respect to cytoplasmic cyclin E staining, suggesting that cytoplasmic cyclin E may be an independent biomarker in these patients.

Tumor lysates from the 152 patients from the Lab00-222 cohort were subjected to both Western blot analysis and IHC with the cyclin E antibody. Nuclear and cytoplasmic staining of cyclin E were quantified and compared with results from the Western blot analysis. Tumors with undetectable or low levels of LMW-E in the Western blot analysis showed predominantly nuclear cyclin E staining, whereas tumors with high levels of LMW-E in the Western blot analysis demonstrated intense, homogeneous cytoplasmic cyclin E staining (Fig. 1A). Logistic regression analyses were performed to determine whether LMW-E and full-length cyclin E expression in the Western blot analysis were associated with cytoplasmic and nuclear staining for cyclin E as a comprehensive score (% positive nuclei × staining intensity). Full-length cyclin E expression in Western blot analysis was an independent predictor of nuclear cyclin E staining, and LMW-E expression in Western blot analysis was an independent predictor of cytoplasmic cyclin E staining (Fig. 1A). Thus, assessment of nuclear and cytoplasmic cyclin E staining in IHC allowed separation into LMW-E (cytoplasmic)–negative and LMW-E–positive groups (Fig. 1B and Supplementary Figs. S1 and S2).

Median follow-up for the Lab00-222 cohort was 9.8 years [95% confidence interval (CI), 9.4–10.5 years], 17.9 years (95% CI, 17.4–18.6 years) for the MDA cohort, 14.8 years (95% CI, 14.3–15.4 years) for the NCI cohort, and 6.3 years (95% CI, 6.1–7.1 years) for the UK cohort. In the combined cohort, 981 patients (39.3%) had negative cytoplasmic cyclin E staining and 1,513 (60.7%) had positive cytoplasmic cyclin E staining, and the distributions did not differ across the four cohorts (Table 1).

In the univariable Cox proportional hazards model, the association between cytoplasmic cyclin E staining and FFR was significant (P < 0.01) for all four cohorts (Table 2). The 5-year FFR rates for the positive cytoplasmic cyclin E staining groups compared with the negative cytoplasmic cyclin E staining groups were 86% compared with 97% for the Lab00-222 cohort, 74% compared with 93% for the MDA cohort, 69% compared with 93% for the NCI cohort, and 64% compared with 92% for the UK cohort (Supplementary Table S3). The HR for a patient with positive cytoplasmic cyclin E staining compared with a patient with negative cytoplasmic cyclin E staining was 4.16 for the Lab00-222 cohort, 3.19 for the MDA cohort, 4.16 for the NCI cohort, and 5.01 for the UK cohort (Table 2).

Kaplan–Meier FFR plots for all four cohorts and all three subtype combinations as a function of cytoplasmic cyclin E staining are shown in Fig. 2. Positive cytoplasmic cyclin E staining was strongly associated with decreased FFR in all four cohorts and across all three subtype combinations, except for HER2-positive and triple-negative subtypes in the Lab00-222 cohort. The associations between cytoplasmic cyclin E staining and FFR were most striking in the hormone receptor–positive subtype.

Clinical and treatment-related factors were represented by the following covariates: age, T stage, nodal status, tumor grade, subtype, chemotherapy, radiotherapy, and endocrine therapy. The multivariable Cox proportional hazards model was fitted in a dataset corresponding to each of the four cohorts. For these analyses, all eight clinical and treatment-related factors and cyclin E staining were initially included in the model as potential risk factors. We sequentially eliminated factors that provided decreased AIC values to obtain the final models. The results of the multivariable analysis are displayed in Table 3, and the corresponding nomograms are displayed in Fig. 3A–D.

Cytoplasmic cyclin E staining was selected in all the four cohorts, and positive cytoplasmic cyclin E staining was strongly associated with poor outcome: FFR HRs were 3.19 (P = 0.012) in the Lab00-222 cohort, 2.48 (P < 0.0001) in the MDA cohort, 3.26 (P < 0.0001) in the NCI cohort, and 6.58 (P < 0.0001) in the UK cohort. In the Lab00-222 cohort, T stage, nodal status, subtype, and chemotherapy were selected along with cytoplasmic cyclin E staining, and T stage was the most significant predictor for FFR (Table 3; Fig. 3A). In the MDA cohort, although age, T stage, nodal status, chemotherapy, and endocrine therapy were selected along with cytoplasmic cyclin E staining, the effect of cytoplasmic cyclin E staining was the strongest among these factors (Table 3; Fig. 3B). In the NCI cohort, T stage, nodal status, tumor grade, and chemotherapy were selected along with cytoplasmic cyclin E staining; cytoplasmic cyclin E staining had the strongest effect on FFR (Table 3; Fig. 3C). Finally, in the UK cohort, age, T stage, nodal status, and endocrine therapy were included in the final model; cytoplasmic cyclin E staining had the strongest effect on FFR (Table 3; Fig. 3D). Figure 3F displays concordance indices for the final models (red) and the final models without cytoplasmic cyclin E staining for all four cohorts. The concordance indices showed substantial gains when cytoplasmic cyclin E staining was added to the model for all cohorts, especially the UK cohort, which had the largest gain.

Next, we sought to determine whether cyclin E had similar prognostic power using the clinical model with Ki67. Information about Ki67 was available for 704 patients in the MDA cohort. Median Ki67 expression was 15% and ranged from 0% to 90%. According to the univariable Cox proportional hazards model, Ki67 was not significant, with a HR for a one-unit change of 1.005 (95% CI, 0.998–1.013). In the multivariable analysis, the backward elimination procedure started from the model, including Ki67, cytoplasmic cyclin E staining, and other clinical factors, including age, T stage, nodal status, tumor grade, subtype, chemotherapy, radiotherapy, and endocrine therapy. The final model included both Ki67 and cytoplasmic cyclin E with other clinical factors, age, T stage, nodal status, chemotherapy, and endocrine therapy. Supplementary Table S4 shows the results of the final model and the corresponding nomogram is displayed in Fig. 3E. The HR of positive cytoplasmic cyclin E staining compared with negative cytoplasmic cyclin E staining was 2.67 (P < 0.0001), and Ki67 was not significant at significance level 0.05 (Supplementary Table S4). Cytoplasmic cyclin E staining outperformed Ki67 and all other clinical factors with the highest LRT statistics with P < 0.0001 (Supplementary Table S3). The nomogram shown in Fig. 3E demonstrates that the prognostic power of cyclin E remained the same as in the model without Ki67 in Fig. 3B for the MDA cohort.

In this study, we examined the relationship between survival outcomes and LMW-E as measured by Western blot analysis and IHC in patients with breast cancer. We found that LMW-E expression in Western blot analysis correlated with cytoplasmic cyclin E staining in IHC. Cytoplasmic cyclin E staining was associated with reduced breast cancer FFR and overall survival in multivariable analysis in four different patient cohorts and across all breast cancer subtypes. In our study, patients with breast cancer whose tumors had no cytoplasmic cyclin E staining had an overall favorable prognosis, and those with any cytoplasmic cyclin E staining had a poor prognosis.

A number of investigators have examined the prognostic value of cyclin E in breast cancer. A meta-analysis of 2,534 patients from 12 studies demonstrated that cyclin E overexpression was associated with a 2.32-fold increased risk of recurrence in univariate analysis and a 1.72-fold increased risk of recurrence in multivariable analysis (40). However, an attempt to validate the association between cyclin E overexpression and poor survival in patients enrolled in the Southwest Oncology Group 9313 trial using IHC for cyclin E on formalin-fixed, paraffin-embedded tissue did not reveal significantly worse outcomes with overexpression of cyclin E (27). We previously reported that LMW-E, not full-length cyclin E, is most active in phosphorylating substrates and that LMW-E has a higher affinity than full-length cyclin E for binding CDK (41). LMW-E is more tumorigenic in transgenic mouse models, and patients whose tumors expressed LMW-E, as measured by Western blot analysis, were shown to be at the highest risk for recurrence and death due to breast cancer (24, 26, 42, 43). Because LMW-E lacks the nuclear localization signal of full-length cyclin E, we previously investigated the subcellular localization of LMW-E and reported that LMW-E accumulates in the cytoplasm, where it binds to CDK2 and retains kinase activity (18). This difference in localization could explain why studies have shown disparate results with respect to cyclin E overexpression and survival.

Key issues in the evaluation of tumor markers for clinical use are utility, magnitude, and reproducibility (44). The clinical utility of cyclin E as a novel tumor marker is multifaceted. In terms of reproducibility, we found that cyclin E expression was stable across multiple cohorts and subtypes of breast cancer. Hence, the expression of cytoplasmic cyclin E is not dependent on the state of differentiation of the disease, but instead is related to the aggressiveness of the disease.

In terms of magnitude, cytoplasmic expression of cyclin E outperformed all other biomarkers examined in three of the four cohorts examined (MDA, NCI, and UK), as shown in the stratified multivariate analysis with FFR as the endpoint (Table 3). To provide data for the clinical relevance of cyclin E as a prognostic marker, we interrogated whether cyclin E can outperform grade and Ki67 in the ER⁺ breast cancer patients. We examined grade and the expression of Ki67 when available and compared with cyclin E, and the results revealed that cyclin E can separate the freedom from survival plots much more significantly than Ki67. In addition, cyclin E assessment can be readily performed in most pathology laboratories and readily interpreted by pathologists without the need for complex algorithms required for interpretation of multigene assays currently in use. The 21-gene recurrence score assay (Oncotype Dx, Genomic Health, Inc.) provides prognostic information for patients with early-stage, ER⁺ breast cancer (45). It also provides predictive information with respect to chemotherapy benefit in patients with a high recurrence score (46). Another multigene assay is the MammaPrint test (Agendia, Inc.), which measures expression of 70 genes and categorizes tumors in patients with early-stage breast cancer into low-risk and high-risk groups (47). This test is similar to the 21-gene recurrence score assay in that it provides both prognostic and predictive information but is applicable to both ER-positive and -negative tumor types (47). IHC for localization of cyclin E provides prognostic information across all subtypes in both early-stage and advanced stage breast cancer. In terms of biomarker utility, cytoplasmic cyclin E can be readily detected with IHC. The same assay will allow detection and scoring of both nuclear and cytoplasmic cyclin E using a readily available antibody and standardized scoring. The results can be interpreted without requiring complex bioinformatics, and it can identify patients who may benefit from alternative treatment strategies targeting cyclin E.

LMW-E has greater affinity for CDK2 than does full-length cyclin E, and LMW-E is a target for CDK2 inhibitors. In HER2-positive breast cancer cells expressing LMW-E, the combination of targeting HER2 with trastuzumab and targeting LMW-E with roscovitine led to synergistic killing (48). We propose that LMW-E not only is a prognostic marker but also could be useful as a predictive marker to identify patients who may benefit from LMW-E–targeted therapy. Previous studies of CDK inhibitors have not shown significant response rates, and therefore, such inhibitors have been presumed to be ineffective. Measurement of cytoplasmic LMW-E may provide the best assessment of patients who could be treated effectively with LMW-E–targeted therapy. In this study, patients with all breast cancer subtypes, including triple-negative breast cancer, had reduced recurrence-free survival when cytoplasmic cyclin E staining was detected. The lack of specific targets for triple-negative breast cancer is a major challenge, and LMW-E could be a target for this aggressive disease.

The function of cytoplasmic cyclin E goes beyond cell proliferation. Historically, the function of nuclear cyclin E has been attributed to cell-cycle progression, and overexpression of nuclear cyclin E leads to deregulation of cell proliferation. Cytoplasmic cyclin E, however, has alternate functions that can affect signal transduction (49), stemness (20, 50), and metabolism (51), to list a few. In fact, comparing cytoplasmic cyclin E with Ki67 revealed that cyclin E was a more significant prognostic indicator than just increased cell proliferation (Supplementary Table S4; Fig. 3E and F). The non–cell-cycle functions of cytoplasmic cyclin E also provide novel avenues for targeted treatment strategies for patients whose tumors express this protein. For example, we recently identified ATP-citrate lyase (ACLY) as a novel interacting protein of LMW-E in the cytoplasm. LMW-E upregulates ACLY enzymatic activity, and ACLY is required for LMW-E–mediated transformation, migration, and invasion in vitro, as well as tumor growth in vivo (51). These studies suggest a novel interplay between LMW-E and ACLY and provide an unexpected link between metabolic pathways and the cell cycle in breast cancer. Therefore, inhibition of ACLY and reduction of lipid accumulation may prove to be beneficial in targeting those breast cancers that display LMW-E expression.

In summary, cytoplasmic cyclin E expression can be readily assessed at diagnosis and is an important marker of prognosis and a target for therapy in breast cancer. Our finding that LMW-E as measured by IHC is associated with poor outcomes provides a rationale for investigating treatment strategies that could specifically target tumors with LMW-E.

K.K. Hunt is a consultant/advisory board member for Armada Health. G.N. Hortobogyi is a consultant/advisory board member for Eli Lilly and Novartis. No potential conflicts of interest were disclosed by the other authors.

Conception and design: K.K. Hunt, M. Yi, K. Keyomarsi

Development of methodology: K.K. Hunt, C. Karakas, A. Biernacka, M. Yi, O. Adjapong, P.A. Thompson, S. Bacus, K.-A. Do, K. Keyomarsi

Acquisition of data (acquired and managed patients, provided facilities, etc.): K.K. Hunt, C. Karakas, A. Biernacka, M. Yi, A.A. Sahin, O. Adjapong, G.N. Hortobagyi, P.A. Thompson, K.L. Cheung, I.O. Ellis, K. Keyomarsi

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): K.K. Hunt, C. Karakas, M.J. Ha, M. Yi, G.N. Hortobagyi, P.A. Thompson, K.L. Cheung, K.-A. Do, K. Keyomarsi

Writing, review, and/or revision of the manuscript: K.K. Hunt, C. Karakas, M.J. Ha, A. Biernacka, G.N. Hortobagyi, M.L. Bondy, K.L. Cheung, I.O. Ellis, S. Bacus, W.F. Symmans, K.-A. Do, K. Keyomarsi

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): K.K. Hunt, M.J. Ha, G.N. Hortobagyi

Study supervision: K.K. Hunt, K. Keyomarsi

We are grateful to Michael Gilcrease, MD, and Thomas Buchholz, MD, for their insightful discussions during the course of this work, Dr. Hannah Wingate, Ms. Dana Richardson and Mr. Tuyen Bui for technical assistance, Debra Simmons, MS, for the preparation of unstained slides, and Stephanie Deming and Erica Goodoff for editing the manuscript.

This work was supported by the NIH through a Cancer Center Support Grant (CA016672; to MD Anderson) and grants CA87548 and CA1522218 (to K. Keyomarsi), as well as through the Susan G. Komen for the Cure grant KG100521 (to K.K. Hunt).

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.