Cancer Investigation, 25:19–26, 2007ISSN: 0735-7907 print / 1532-4192 onlineCopyright c© Informa HealthcareDOI: 10.1080/07357900601128825

ORIGINAL ARTICLE

Quantitative Analysis of Breast Cancer TissueMicroarrays Shows High Cox-2 Expression Is

Associated with Poor OutcomeMaciej P. Zerkowski, B.S.,1 Robert L. Camp, M.D., Ph.D.,1 Barbara A. Burtness, M.D.,2

David L. Rimm, M.D., Ph.D.,1 and Gina G. Chung, M.D.2

Department of Pathology, Yale University School of Medicine, New Haven, Connecticut, USA.1

Yale Cancer Center and Department of Internal Medicine, Section of Medical Oncology,Yale University School of Medicine, New Haven, Connecticut, USA.2

ABSTRACT

Epidemiologic and preclinical studies suggest that cyclooxygenase-2 (Cox-2) may promotetumor growth and spread by affecting angiogenesis and apoptosis in breast cancer. Using atissue microarray (TMA), we analyzed the expression and subcellular localization of Cox-2 byAQUA and X-tile, our algorithms for quantitative analysis of protein expression and determi-nation of optimal cutpoints. Our TMA consisted of 669 Stage I–III primary breast cancers. Thetotal tumor and subcellular expression of Cox-2 were then correlated with clinicopathologicfactors and with survival. Cox-2 expression appeared higher in malignant than in benign tissueand was predominantly membrane/cytoplasmic (i.e. non-nuclear). X-tile determines an optimumcutpoint on a training set then uses this cutpoint on a validation set. This cutpoint was 19.3(top 44 percent defined as positive) with high nonnuclear Cox-2 expressers having significantlyworse survival. Cox-2 expression also was inversely associated with estrogen receptor (ER) andprogesterone receptor (PR), and directly associated with nuclear grade. Multivariate analysisshowed that Cox-2 remained a significant prognostic factor for survival independent of tumorsize, nodal status, ER, Her2/neu, and grade. In summary, Cox-2 is overexpressed in breast neo-plasms, is associated with other markers of poor prognosis, and is significantly associatedwith worse survival independent of known prognostic factors. Furthermore, AQUA and X-tileanalysis suggest an optimal cutpoint that may be helpful in future investigations of Cox-2 andspecifically, in studies looking at its expression as a predictive biomarker in clinical trials ofCox-2 inhibitors in breast cancer.

This work was supported in part by a grant from the Anna & ArgallHull/Yale Cancer Center Translational Research Award and theGreenwich Breast Cancer Alliance Young Investigator Awardawarded to G.G. Chung.The work in this article is original and has not been previouslypresented or published.Keywords: Cox-2, Breast cancer, Quantitative analysis, Tissuemicroarrays, Prognosis.Correspondence to:Gina G. Chung, M.D.Yale Cancer CenterSection of Medical Oncology333 Cedar St., P.O. Box 208032New Haven, CT 06520e-mail: gina.chung@yale.edu

INTRODUCTION

Two isoforms of Cox have been described (Cox-1 and Cox-2).Cox-1 is mostly expressed constitutively and is involved pre-dominantly in physiologic functions. Cox-2 is inducible in re-sponse to a variety of stimuli. Both enzymes are involved inthe production of prostaglandins and thromboxanes from freearachidonic acid.

More recently, these enzymes, in particular Cox-2 and itsderived prostaglandin PGE2, have been linked to carcinogen-esis. The first suggestion of this relationship in breast cancerstemmed from observations that PGs and PGE2 levels were el-evated in breast tumors and in metastases (1). Subsequently,numerous epidemiologic studies have shown an associationbetween reduced colon cancer risk and use of nonsteroidalanti-inflammatory drugs (NSAIDS) which are nonspecific Cox

19

Can

cer

Inve

st D

ownl

oade

d fr

om in

form

ahea

lthca

re.c

om b

y M

erce

r U

nive

rsity

on

10/2

9/14

For

pers

onal

use

onl

y.

inhibitors (2, 3). There are far fewer studies looking at the asso-ciation between Cox inhibitors and breast cancer risk. However,several studies, including a recent meta-analysis, found a com-bined relative risk of 0.8 (4).

Several studies in breast cancer have shown overexpressionof Cox-2 mRNA and/or protein by various methods. In addi-tion, some of these studies have shown an association with poorprognostic factors and worse outcome, whereas others have notconfirmed these findings (5, 6). The majority of these studies,however, have been performed on relatively small cohorts ofpatients with short follow-up.

Functionally, Cox-2 has been associated with the develop-ment, growth, and spread of cancers by stimulation of angio-genesis, inhibition of apoptosis, and enhancement of invasive-ness. Molecular evidence for the role of Cox-2 came first inanimal models of colorectal cancer showing that disruption ofthe Cox-2 gene resulted in decreased tumor formation and thatselective and nonselective Cox inhibitors also had anti-tumoreffects (7, 8). Similarly, transgenic mice with murine mammarytumor virus (MMTV) promoter driven Cox-2 expression devel-oped spontaneous breast carcinomas and again Cox inhibitors in-hibited the development and progression of these tumors (9, 10).These preclinical studies, thus, suggest that the forced expres-sion of Cox-2 in these models is sufficient for tumorigenesis andthat Cox-inhibitors are both chemopreventive and therapeutic.

Based on these preclinical data, numerous clinical trials wereinitiated with selective Cox-2 inhibitors in the prevention andtreatment of breast cancer. The prognostic value of Cox-2, how-ever, is not well established, with only limited data that can linkthe expression to meaningful clinical outcome measures. In ad-dition, previous in situ measurements have graded on nominalmanual scoring systems limited by subjectivity and difficultiesin reproducibility. Finally, there are essentially no data on thepredictive value of Cox-2 expression. Current studies of Cox-2inhibitors do not have any prospective requirements for Cox-2expression; however, a correlation between this and outcomeand response would be highly desirable. Furthermore, given re-cent data showing increased cardiovascular risks associated withCox-2 inhibition, a closer re-examination and validation of thisenzyme as a prognostic marker and as a therapeutic target iswarranted to more accurately assess the risk: benefit ratio. Be-cause protein expression is a continuous variable and cutpointsfor positive versus negative expression is often arbitrarily de-termined, we used our automated analysis of protein in tissuecalled AQUA (11). Together with a novel method of determin-ing optimum cutpoints of continuous data based on outcomeinformation called X-tile (12), we determined the prognosticvalue of Cox-2 expression in a large cohort breast cancer tissuemicroarray.

MATERIALS AND METHODS

Tissue microarray

Our cohort consisted of 669 formalin-fixed, paraffin-embedded blocks of primary breast cancer specimens obtained

Table 1. Patient cohort characteristics (n = 669)

N (%) Median (Range)

Age (years) 58 (24–88)< 50 178 (27)≥ 50 475 (71)Not specified 16 (2)Histology

Infiltrating duct 550 (84)Infiltrating lobular 13 (2)Carcinoma NOS 86 (13)Other 7 (1)

Tumor Size (cm) 2.5 (0.13–20)<2 cm 215 (32)≥ 2 cm, <5 cm 286 (43)≥ 5 cm 100 (15)Not specified 68 (10)Nodal status

Positive 324 (49)Negative 331 (51)

Nuclear grade1 113 (19)2 317 (53)3 171 (28)

ERPositive 331 (53)Negative 299 (47)

PRPositive 305 (50)Negative 303 (50)

HER20 386 (62)1 123 (20)2 41 (7)3 68 (11)

Follow-up (years) 8.9 (0.17–53)

from 1953 to 1983 (Table 1). The cases were nearly evenly di-vided between node positive and node negative cases with amedian follow up time of 8.9 years. Although full treatment in-formation was unavailable for this cohort, most patients weretreated with a combination of surgical excision +/− local ir-radiation and/or hormonal therapy. Approximately 10 percentof patients received some form of chemotherapy. Although notexplicitly stated, it appears that both chemotherapy and radia-tion were delivered postoperatively in the majority of cases. Thesamples used for the TMA were primary surgical specimens,that is, pretreatment tumor samples. This is important as Cox-2expression has been shown to be affected by chemoradiation.The blocks were retrieved from the archives of the Yale Uni-versity Department of Pathology in accordance with the localInstitutional Review Board. Representative areas of invasive tu-mor were identified (R.L.C and D.L.R.) and 0.6-mm diametercores were placed into a recipient block using a precision array-ing instrument (Beecher Instruments, Silver Spring, MD, USA).Five µm sections were affixed to adhesive slides using a UVcross-linkable tape transfer system (Instrumedics Inc., Hacken-sack, NJ, USA), then coated in paraffin and stored in a nitro-gen chamber prior to staining to prevent antigen oxidation anddegeneration (13).

20 M. P. Zerkowski et al.

Can

cer

Inve

st D

ownl

oade

d fr

om in

form

ahea

lthca

re.c

om b

y M

erce

r U

nive

rsity

on

10/2

9/14

For

pers

onal

use

onl

y.

Immunohistochemistry

Staining tissue microarray slides for AQUA has been pre-viously described (11). Briefly, slides were deparaffinized inxylene, rinsed in ethanol, and rehydrated. Antigen retrieval wasperformed by pressure cooking for 15 minutes in 6.5 mM sodiumcitrate buffer. Endogenous peroxidase was quenched by im-mersing the array in a 2.5 percent methanol/hydrogen perox-ide buffer for 30 minutes. Nonspecific background staining wasfurther minimized by preincubating the array with 0.3 percentbovine serum albumin in 0.1 M tris-buffered saline (pH 8.0) for1 hour. Cox-2 expression was detected using a mouse mono-clonal antibody at a 1:250 dilution (Cayman Chem., Ann Arbor,MI, USA). For purposes of our automated analysis, tumor cellswere also differentiated from stroma with a cytokeratin rab-bit polyclonal antibody (DAKO, Carpinteria, CA, USA). Thisprimary antibody cocktail was incubated overnight at 4◦C ina humidity chamber. Goat anti-mouse antibody conjugated toa horseradish peroxidase-decorated dextran polymer backbone(Envision; DAKO, Carpinteria, CA, USA) was used as a sec-ondary reagent to detect the bound primary Cox-2 antibody andCy5-tyramide was used to visualize the amplified signal. The cy-tokeratin was visualized with a Cy3-conjugated secondary anti-body and the array was then counterstained with 4′,6-diamidino-2-phenylindole (DAPI) to localize nuclei.

Automated image acquisition and analysis

Image acquisition and automated analysis have also beendescribed extensively in our previous work (11). In brief,slides are scanned using a customized, computer-controlled epi-fluorescence microscope (Olympus BX-51, with xy-stage andz controller). Images of each spot at different wavelengths areautomatically acquired with a high-resolution monochromaticcamera (CookePCO, Romulus, MI, USA). Using this stack ofuncompressed images, the AQUA software then allows one todistinguish between areas of tumor and stromal elements us-ing the cytokeratin stain, resulting in an unique binary cytok-eratin tumor mask for each spot. Furthermore, the cytokeratinand DAPI stains are used to assign each pixel under the tumormask into nonoverlapping membrane, cytoplasmic and nuclearlocales. AQUA scores/units for Cox-2 are then calculated thatcorrespond to the average signal intensity divided by locale area.These AQUA scores typically range between 0 and 255, althoughwe have found that above approximately 150 the images areoversaturated. This information can then be exported in a formatsuitable for analysis by standard software packages or by X-tile.

Statistical analysis

All analyses were performed using Statview software (ver-sion 5.0.1; SAS Institute Inc., Cary, NC, USA) and X-tile. TheCox proportional-hazards model was used to perform multivari-ate analysis to determine relative risk and independent signifi-cance. Correlations between markers and with clinicopathologicparameters were determined using chi-square analysis. Survivalcurves were calculated using Kaplan-Meier analysis with assess-

ment of statistical significance using the Mantel-Cox log-ranktest. X-tile plots provide an overview of every possible way ofdividing a population into low/high or low/medium/high-levelmarker expression and associations are calculated at each divi-sion by standard tests, including the log-rank test for survival.However, since it is statistically invalid to test multiple divi-sions and accept the best p-value, rigorous statistical evaluationis achieved by defining the optimal division by outcome in a“training set” and then validating in a separate patient cohort(“validation set”).

RESULTS

Immunostaining patterns

Our staining using epifluoresence showed a predominantmembrane/cytoplasmic distribution in the tumor cells (Figure 1).This is similar to descriptions of immunostaining patterns pre-viously described in the literature (5). Our breast cancer TMAalso included 40 benign breast tissue controls which were benignareas of epithelium from patients who had surgical excision forbreast cancer. Although Cox-2 staining on these samples wasof a similar pattern compared with the cancers, they appearedsignificantly less compared with the malignant samples. Statis-tical analysis however, was not feasible because of the smallnumber of these spots. Figure 2 shows frequency histograms ofCox-2 AQUA scores in the different subcellular compartments.Because the expression profile and association with outcome formembrane and cytoplasmic locales appeared similar, we hence-forth used a combined non-nuclear locale for descriptions ofCox-2 expression and for subsequent analysis. The heavy verti-cal line in panel A is AQUA score of 19.3 determined by X-tileto be the optimal cutpoint (see below). Nuclear expression wasmuch lower. This is in accordance with what was qualitativelyseen by eye. Furthermore, Cox-2 expression in the nuclear com-partment was not significantly associated with any of the otherprognostic variables, nor was it associated with outcome.

Validation of microarray cohort

Table 1 outlines patient characteristics of our cohort. In orderto validate our cohort of patients, we first assessed several tradi-tional markers of breast cancer. Univariate analysis showed thatlarge tumor size, positive nodal status, low ER levels, and highnuclear grade were all significant predictors of poor prognosis(Table 2). Although high Her2/neu expression assessed by im-munohistochemistry showed a trend toward worse outcome, thisdid not achieve statistical significance. In our cohort, high non-nuclear expression of Cox-2 also was significantly associatedwith poor disease-specific and overall survival when analyzedas a continuous variable or as a binary variable using the opti-mized cutpoint determined by X-tile analysis (see below).

Cox-2 expression and outcome

AQUA scores for nonnuclear expression of Cox-2 rangedbetween 4.7 and 120.4. We initially divided the patients into

Cox-2 Expression in Breast Cancer 21

Can

cer

Inve

st D

ownl

oade

d fr

om in

form

ahea

lthca

re.c

om b

y M

erce

r U

nive

rsity

on

10/2

9/14

For

pers

onal

use

onl

y.

Figure 1. Immunostains of breast cancer with Cox-2. Panels A and B show representative TMA histospots of low and high levels of Cox-2expression, respectively. These images show that Cox-2 is predominantly expressed within the tumor cells. The merged quadrants also showthat when Cox-2 is overexpressed, it is localized by AQUA predominantly to the membrane/cytoplasmic locale (yellow in merged quadrant ofpanel B represents Cox-2 co-localized with cytokeratin in the membrane/cytoplasm compartment).

quartiles. Figure 3 (inset) shows that the top two and bottomtwo quartiles appear to have a natural separation. Therefore, wedivided the entire group into the top and bottom two quartiles forKaplan-Meier survival analysis. This showed that the top twoquartiles (48 percent) had significantly worse disease-specificsurvival (Figure 3).

The decision to divide the scores into quartiles however, isof course somewhat arbitary. We therefore used X-tile to lookat other cutpoints and to validate the relationship of Cox-2 tooutcome. Although the X-tile “heat maps” (Figure 4, panels Aand D) show a relatively diffuse pattern suggestive of a linearcorrelation with outcome (i.e., high Cox-2 expression is alwaysassociated with a poorer outcome no matter where the divisionis made), for the purposes of clinical studies, we searched for anoptimal cutpoint. In order to rigorously validate the statisticalsignificance of a cutpoint, X-tile divides the entire cohort intotwo populations matched for survival, one designated as thetraining set and the other as the validation set.

This evaluation in the training set determined an optimal cut-point of 19.3 (top 44 percent of Cox-2 expressers) with respectto disease-specific survival and overall survival (Figure 4, panelsB and E). When this cutpoint was tested in the validation set, theP-value was 0.0189 and 0.0055 for disease-specific and overallsurvival respectively (Figure 4, panels C and F). This is highlysignificant in light of the fact that the validation set was now halfthe size of the original cohort.

Association with other biomarkers

We then performed Chi square analysis to determine pos-sible associations between Cox-2 expression and other mark-

ers (Table 3). This showed that Cox-2 positivity (as defined byX-tile) was significantly associated with low ER and PR expres-sion and with high nuclear grade. In our cohort, there was no sig-nificant correlation with Her2/neu as has been shown in severalprior studies. This may be a reflection of the immunohistochem-ical evaluation for Her2/neu that was employed for this analysis.

Multivariate analysis

To assess the independent prognostic significance of Cox-2and the other variables assessed, we performed multivariate anal-ysis (Table 2). This showed that tumor size and nodal status werethe two most significant variables as expected, but that high Cox-2 expression (using the cutpoint of 19.3) also was a significantvariable with a relative risk of 1.66, similar to that seen for tumorsize.

DISCUSSION

Much concern has been raised recently due to studies ofCox-2 inhibitors primarily in the preventative setting of col-orectal cancer that have shown increased cardiovascular risksassociated with the use of these drugs. There are still other stud-ies completed and ongoing that did not have similar findings.Due to these concerns, a number of studies in both adjuvant andadvanced breast cancer of Cox-2 inhibitors has been suspended.These results highlight the need for closer examinations of thetoxicities of targeted therapeutics but also to more selectivelychoose patients for clinical studies. Thus, a readily measurableand reproducible predictive biomarker for Cox-2 inhibitors maybetter help to identify subsets of patients with advanced cancer

22 M. P. Zerkowski et al.

Can

cer

Inve

st D

ownl

oade

d fr

om in

form

ahea

lthca

re.c

om b

y M

erce

r U

nive

rsity

on

10/2

9/14

For

pers

onal

use

onl

y.

Figure 2. Frequency histograms of Cox-2 expression. Panel Ashows nonnuclear (i.e., membrane/cytoplasmic) levels of expres-sion of Cox-2 and panel B shows nuclear expression. The y-axis(count) represents the number of cases in the TMA with AQUAscores in the range specified on the x-axis. The heavy verticalline in panel A represents the optimal cutpoint of 19.3 as de-termined by X-tile. Nonnuclear expression was more widely dis-tributed, whereas nuclear expression was more tightly skewed to-ward the lowest AQUA scores. The mean value for nonnuclearexpression was 23.8 versus 12.5 for nuclear expression.

Table 2. Univariate analysis

Marker P-value RR (95% CI)

ER positive <0.0001 0.59 (0.45–0.77)HER2 positive 0.077 1.33 (0.97–1.83)Tumor size >2 (cm) <0.0001 1.99 (1.51–2.62)Nodal status positive <0.0001 2.77 (2.11–3.64)Nuclear grade high 0.0001 1.69 (1.29–2.23)Cox-2 AQUA 0.0007 1.68 (1.25–2.26)

nonnuclear positive (>19.3)Cox-2 AQUA 0.0078

nonnuclear positive (continuous)

Multivariate analysisMarker P-value RR (95% CI)

ER positive 0.0554 0.72 (0.52–1.00)HER2 positive 0.0374 1.51 (1.02–2.23)Tumor size >2 (cm) 0.0022 1.68 (1.20–2.33)Nodal status positive <0.0001 2.39 (1.67–3.45)Nuclear grade high 0.9283 0.98 (0.69–1.40)Cox-2 AQUA 0.0084 1.66 (1.12–2.16)

nonnuclear positive (>19.3)

Figure 3. Kaplan-Meier survival curves of Cox-2 quartiles. Thetop two quartiles (top 52 percent) of Cox-2 expressers had a sig-nificantly better 10 year disease specific survival (P = 0.0155)compared with the bottom two quartiles (bottom 48 percent). Asimilar result also was seen for overall survival at 10 years (datanot shown). Inset shows Kaplan-Meier survival curves for all fourquartiles.

who may have maximal benefit while limiting exposure andtoxicity to those with little likelihood of benefit.

We used our recently developed automated imaging systemusing molecular colocalization techniques to quantitatively as-sess the prognostic significance of Cox-2 in our breast cancertissue microarray. Our results showed that high non-nuclear lev-els of Cox-2 within the tumor significantly correlated with tradi-tional markers of poor outcome (e.g., low ER, high grade), wassignificantly associated with worse disease-specific and over-all survival, and that this was independent of other traditionalprognostic markers in breast cancer.

Several prior studies have looked at the immunohistochem-ical expression of Cox-2 in breast cancer, some with a stainingindex which combines intensities with the percentage of pos-itive cancer cells. Although this gives some crude measure ofquantification, it is still subject to much observer interpretation.Several of these studies have shown an association of Cox-2with markers of poor prognosis, however there is little unifor-mity between Cox-2 expression and with survival. Moreover,the majority of these studies have not shown an association withoverall survival and none has shown significance on multivariateanalysis. A study by Wulfing et al., analyzed Cox-2 expressionin a TMA of 200 patients with invasive breast cancer (14).Using a semi-quantitative scoring system (0–3), they foundCox-2 positivity (defined as a score of 2–3) in 40.6 percent ofsamples. In their analyses, Cox-2 was directly correlated withtumor size, lymphovascular invasion, and histologic grade andinversely correlated with PR and axillary lymph node status.There was no significant association between Cox-2 levels andsurvival. Similar findings were also presented by Lim et al. ina smaller cohort (15). In contrast, using an immunoreactivityscore derived from percent staining and intensity as a measureof Cox-2 expression, Spizzo et al. found 48.6 percent Cox-2positivity (16). Cox-2 expression was not associated with any

Cox-2 Expression in Breast Cancer 23

Can

cer

Inve

st D

ownl

oade

d fr

om in

form

ahea

lthca

re.c

om b

y M

erce

r U

nive

rsity

on

10/2

9/14

For

pers

onal

use

onl

y.

Figure 4. X-tile map and survival curves for Cox-2 expression. Panels A and D show X-tile “heat maps” of Cox-2 expression for DoD (diseasespecific survival-panel A) and for alive/dead (overall survival-panel D). In the heat maps, each point (pixel) represents the data from a givenset of divisions into three groups in the triangular grid and two-way divisions in the horizontal strip. The vertical axis represents all possible“high” populations, with the size of the high population increasing from top to bottom. Similarly, the horizontal axis represents all possible “low”populations, with the size of the low population increasing from left to right. Data along the hypotenuse represent divisions in which all casesbelong to either the high or low population. Specifically, a chi squared value is calculated for every possible division of the population shown onthe grid using a color code. Indirect associations between marker expression and survival (e.g., high expression connotes poorer survival) arecolored red, whereas direct associations are colored green. Coloration of the plot represents the strength of the association at each division,ranging from low (dark, black) to high (bright, green or red). The coloring was fairly diffuse in all the groups, although it appeared to cluster nearthe median. The cutpoint of 19.25 shown in the 2-way division on the bottom strip divided the cohort into top 44 percent and bottom 56 percentof expression. Panels B, C, E, and F are Kaplan-Meier survival curves of Cox-2 at this cutpoint for training and validation sets with respect toboth 10 year disease-specific and overall survival.

of the traditional prognostic categories (e.g., tumor size, nodalstatus, ER/PR status), but was associated with worse disease-free and overall survival. Ristimaki et al., analyzed Cox-2expression in a large cohort (5). They also used a combinationof percentage and intensity to score the immunostains resultingin a 37.4 percent positive rate. Similar to the study by Wulfing

Table 3. Chi square Analysis

Marker ER PR HER2 Tumor size >2 cm Nodal status positive High nuclear grade

ERPR <0.0001∗

HER2 <0.0001∗† 0.0057∗†

Tumor Size >2 cm 0.4899 0.2582 0.9046Nodal Status Positive <0.0001∗† <0.0001∗† 0.0878 0.0015∗

High Nuclear Grade 0.0003∗† <0.0001∗† <0.0001∗ 0.0906 <0.0001∗

Cox-2 AQUA Non-Nuc. pos. 0.0355∗† 0.0097∗† 0.1525 0.1815 0.4096 <0.0001∗

∗P-value <0.05.†Inverse correlation (corrections for multiple comparisons were not made).

et al., Cox-2 was significantly correlated with tumor size andgrade and inversely correlated with hormone receptor status andnodal status. However, this study showed a strong correlationbetween elevated Cox-2 expression and worse distant disease-free survival, although multivariate analysis did not confirmCox-2 as an independent prognostic factor. Finally, Denkert

24 M. P. Zerkowski et al.

Can

cer

Inve

st D

ownl

oade

d fr

om in

form

ahea

lthca

re.c

om b

y M

erce

r U

nive

rsity

on

10/2

9/14

For

pers

onal

use

onl

y.

et al., recently reported an association of Cox-2 expression on acohort of 221 patients again using a combination of percentageand intensity and showed association with markers of poorprognosis but no significance on multivariate analysis (17).

Whereas our results confirm some of the findings fromthe studies above, they also show significant association withdisease-specific and overall survival independent of other prog-nostic variables in multivariate analysis. In addition, althoughcytoplasmic staining usually has been noted to be the predom-inant pattern in prior studies, we also detected moderate lev-els in the membrane. Because it did not give any significantprognostic information distinct from cytoplasmic expression,we chose to look at the combined membrane/cytoplasmic ex-pression (i.e., nonnuclear) in further analysis. Although severalcases appeared to have some nuclear expression, quantitativeanalysis with AQUA clearly shows in Figure 2 that the nuclearlevels were overall much lower and most likely represented back-ground. Moreover, nuclear expression did not have any signifi-cant association with other prognostic markers or with survival.

A limitation of our study, as is often the case in any retrospec-tive analysis, was the lack of complete treatment information onour cohort of patients. Clearly, surgical, radiotherapy, and sys-temic treatments varied over the period of time on our array.However, the vast majority (88 percent) of the patients werediagnosed and treated between 1960 and 1980. In that time, sys-temic chemotherapies were not often utilized, and, when used,often only for more advanced node positive disease. There wasno significant difference in our results looking at only node pos-itive or node negative subsets, suggesting although certainly notproving, that these results may be chemotherapy treatment in-dependent. The vast majority of patients received only surgeryand/or radiation and in this era and there was fairly minimal useof tamoxifen.

It is interesting to note that although prior studies have useddifferent methodologies, most studies have shown relativelysimilar percentages of “positive” Cox-2 cases (37–49 percent).Our histograms (Figure 2) clearly show a multimodal distribu-tion which would most likely have been lost on a manual scoringsystem. X-tile provides a means of dividing a population in everypossible way into 2 or 3 levels of marker expression. The X-tilesoftware also provides a method of dividing a single cohort intotraining and validation subsets for p-value estimation when sep-arate training and validation cohorts are not available. Althoughsubjective assessments of frequency histograms may be helpfulin giving insights into the biological significance of a marker (asit may be for Cox-2), this is often unreliable and even can bemisleading. The X-tile graph potentially provides more accurateinsight into the biologic nature of a marker. For Cox-2, X-tiledemonstrates a diffuse distribution of risk that always showsworse survival with high Cox-2 expression and is corroboratedby continuous univariate analysis showing significant associa-tion of Cox-2 with survival. This cutpoint is demonstrated inFigure 2 and shows that whereas it appears to coincide with thelowest cluster of Cox-2 expressers, other clusters of higher ex-pressers are not distinguished. These results suggest complexregulation of Cox-2 expression and indeed is corroborated by

work that have shown multiple levels of regulation by differentfactors, including Her2/neu and aromatase/estrogen levels (18,19). Nevertheless, for practical purposes, the optimal cutpointfor this “continuous” marker was determined to be 19.3 yieldinghighly significant P-values for both training and validation sets.For Cox-2 in our cohort, this cutpoint as optimized and validatedby X-tile did correlate relatively closely with standard quartileanalysis with division into the highest two quartiles (cutpoint of17.4) versus bottom two. Only rigorous validation in prospectivestudies, however, will ultimately validate the clinical utility ofCox-2 expression as a prognostic and/or predictive biomarker.

In conclusion, using AUQA and X-tile analysis, our studiesshowed that Cox-2 expression was likely upregulated in invasivebreast tumors, and that elevated non-nuclear levels within the tu-mor were significantly associated with more standard markers ofpoor prognosis in breast cancer. There was a fairly linear correla-tion with survival, despite the multimodal appearance of the his-togram. Using our optimal cutpoint analysis, the high expressorshad significantly worse disease-specific and overall survival, in-dependent of tumor size, nodal status, grade, and ER levels.

REFERENCES1. Rolland, P.H.; Martin, P.M.; Jacquemier, J.; Rolland, A.M.; Toga, M.

Prostaglandin in human breast cancer: evidence suggesting thatan elevated prostaglandin production is a marker of high metastaticpotential for neoplastic cells. J. Natl. Cancer Inst. 1980, 64, 1061–1070.

2. Thun, M.J.; Namboodiri, M.M.; Heath, C.W., Jr. Aspirin use andreduced risk of fatal colon cancer. N. Engl. J. Med. 1991, 325,1593–1596.

3. Giovannucci, E.; Rimm, E.B.; Stampfer, M.J.; Colditz, G.A.;Ascherio, A.; Willett, W.C. Aspirin use and the risk for colorec-tal cancer and adenoma in male health professionals. Ann. Intern.Med. 1994, 121, 241–246.

4. Khuder, S.A.; Mutgi, A.B. Breast cancer and NSAID use: a meta-analysis. Br. J. Cancer 2001, 84, 1188–1192.

5. Ristimaki, A.; Sivula, A.; Lundin, J.; et al. Prognostic significanceof elevated cyclooxygenase-2 expression in breast cancer. CancerRes. 2002, 62, 632–635.

6. Half, E.; Tang, X.M.; Gwyn, K.; Sahin, A.; Wathen, K.; Sinicrope,F.A. Cyclooxygenase-2 expression in human breast cancers andadjacent ductal carcinoma in situ. Cancer Res. 2002, 62, 1676–1681.

7. Oshima, M.; Murai, N.; Kargman, S.; et al. Chemoprevention of in-testinal polyposis in the Apcdelta716 mouse by rofecoxib, a specificcyclooxygenase-2 inhibitor. Cancer Res. 2001, 61, 1733–1740.

8. Oshima, M.; Dinchuk, J.E.; Kargman, S.L.; et al. Suppression ofintestinal polyposis in Apc delta716 knockout mice by inhibition ofcyclooxygenase 2 (COX-2). Cell 1996, 87, 803–809.

9. Liu, C.H.; Chang, S.H.; Narko, K.; et al. Overexpression ofcyclooxygenase-2 is sufficient to induce tumorigenesis in trans-genic mice. J. Biol. Chem. 2001, 276, 18563–18569.

10. Chang, S.H.; Liu, C.H.; Conway, R.; et al. Role of prostaglandin E2-dependent angiogenic switch in cyclooxygenase 2-induced breastcancer progression. Proc. Natl. Acad. Sci. USA 2004, 101, 591–596.

11. Camp, R.L.; Chung, G.G.; Rimm, D.L. Automated subcellular local-ization and quantification of protein expression in tissue microar-rays. Nat. Med. 2002, 8, 1323–1327.

12. Camp, R.L.; Dolled-Filhart, M.; Rimm, D.L. X-tile: a new bio-informatics tool for biomarker assessment and outcome-based cut-point optimization. Clin. Cancer Res. 2004, 10, 7252–7259.

Cox-2 Expression in Breast Cancer 25

Can

cer

Inve

st D

ownl

oade

d fr

om in

form

ahea

lthca

re.c

om b

y M

erce

r U

nive

rsity

on

10/2

9/14

For

pers

onal

use

onl

y.

13. DiVito, K.A.; Charette, L.A.; Rimm, D.L.; Camp, R.L. Long-termpreservation of antigenicity on tissue microarrays. Lab. Invest.2004, 84, 1071–1078.

14. Wulfing, P.; Diallo, R.; Muller, C.; et al. Analysis of cyclooxygenase-2 expression in human breast cancer: high throughput tissue mi-croarray analysis. J. Cancer. Res. Clin. Oncol. 2003, 129, 375–382.

15. Lim, S.C. Role of COX-2, VEGF and cyclin D1 in mammary infil-trating duct carcinoma. Oncol. Rep. 2003, 10, 1241–1249.

16. Spizzo, G.; Gastl, G.; Wolf, D.; et al. Correlation of COX-2 andEp-CAM overexpression in human invasive breast cancer and itsimpact on survival. Br. J. Cancer 2003, 88, 574–578.

17. Denkert, C.; Winzer, K.J.; Muller, B.M.; et al. Elevated expressionof cyclooxygenase-2 is a negative prognostic factor for diseasefree survival and overall survival in patients with breast carcinoma.Cancer 2003, 97, 2978–2987.

18. Brueggemeier, R.W.; Richards, J.A.; Petrel, T.A. Aromatase andcyclooxygenases: enzymes in breast cancer. J. Steroid. Biochem.Mol. Biol. 2003, 86, 501–507.

19. Subbaramaiah, K.; Norton, L.; Gerald, W.; Dannenberg, A.J.Cyclooxygenase-2 is overexpressed in HER-2/neu-positive breastcancer: evidence for involvement of AP-1 and PEA3. J. Biol. Chem.2002, 277, 18649–18657.

26 M. P. Zerkowski et al.

Can

cer

Inve

st D

ownl

oade

d fr

om in

form

ahea

lthca

re.c

om b

y M

erce

r U

nive

rsity

on

10/2

9/14

For

pers

onal

use

onl

y.