Gene Expression Profiles in Paraffin-Embedded Core Biopsy Tissue Predict Response to Chemotherapy in Women With Locally Advanced Breast Canc
http://www.100md.com
《临床肿瘤学》
the Istituto Nazionale Tumori, Milan, Italy
The University of Texas M.D. Anderson Cancer Center, Houston, TX: Genomic Health Inc, Redwood City, CA
Millennium Pharmaceuticals Inc, Cambridge, MA
ABSTRACT
PURPOSE: We sought to identify gene expression markers that predict the likelihood of chemotherapy response. We also tested whether chemotherapy response is correlated with the 21-gene Recurrence Score assay that quantifies recurrence risk.
METHODS: Patients with locally advanced breast cancer received neoadjuvant paclitaxel and doxorubicin. RNA was extracted from the pretreatment formalin-fixed paraffin-embedded core biopsies. The expression of 384 genes was quantified using reverse transcriptase polymerase chain reaction and correlated with pathologic complete response (pCR). The performance of genes predicting for pCR was tested in patients from an independent neoadjuvant study where gene expression was obtained using DNA microarrays.
RESULTS: Of 89 assessable patients (mean age, 49.9 years; mean tumor size, 6.4 cm), 11 (12%) had a pCR. Eighty-six genes correlated with pCR (unadjusted P < .05); pCR was more likely with higher expression of proliferation-related genes and immune-related genes, and with lower expression of estrogen receptor (ER) –related genes. In 82 independent patients treated with neoadjuvant paclitaxel and doxorubicin, DNA microarray data were available for 79 of the 86 genes. In univariate analysis, 24 genes correlated with pCR with P < .05 (false discovery, four genes) and 32 genes showed correlation with P < .1 (false discovery, eight genes). The Recurrence Score was positively associated with the likelihood of pCR (P = .005), suggesting that the patients who are at greatest recurrence risk are more likely to have chemotherapy benefit.
CONCLUSION: Quantitative expression of ER-related genes, proliferation genes, and immune-related genes are strong predictors of pCR in women with locally advanced breast cancer receiving neoadjuvant anthracyclines and paclitaxel.
INTRODUCTION
When chemotherapy is administered preoperatively to women with breast cancer, a rate of clinical complete and partial response exceeding 70% is achieved with many regimens.1-3 However, a lower percentage ranging from about 5% to 38% of patients achieve tumor eradication when response is assessed with careful pathologic examination of the surgical specimen after chemotherapy (pathologic complete response [pCR]).1-4 Several studies show that, together with nodal status at surgery, pCR stands out as the most significant independent variable associated with the likelihood of benefit as measured by disease-free and overall survival.4 In principle, diagnostic tests that predict which patients will experience a pCR will identify patients who will likely benefit from treatment, and, will also indentify individuals with low or no likelihood of benefit, sparing them the toxicity of an inactive treatment and allowing earlier application of alternative approaches.
Over the last several years a number of laboratories have demonstrated that DNA microarray–based gene expression profiling (GEP) of patient tumor tissue provides prognostic and predictive information.5 In breast cancer in particular, several DNA microarray studies suggest that GEP can provide better prognostic discrimination than traditional clinical tools, and moreover can predict response to preoperative chemotherapy.6-12 Continued progress with DNA microarrays in this area, toward biomarker discovery, clinical validation, and ultimately routine clinical use, is limited by DNA microarray requirements for fresh tumor tissue.13,14
Recent work shows that real time reverse transcriptase polymerase chain reaction (RT-PCR) can quantify gene expression in formalin-fixed paraffin-embedded (FPE) tumor tissue specimens with high reproducibility, specificity and sensitivity.15,16 Using this technique to measure expression of 250 candidate genes, three studies in a total of 447 patients were carried out to identify a 21-gene panel and Recurrence Score (RS) algorithm. The RS was clinically validated to quantify the risk of distant recurrence in 668 women with estrogen receptor–positive, axillary lymph node-negative breast tumors treated with adjuvant tamoxifen in a clinical trial conducted by the National Surgical Adjuvant Breast and Bowel Project (NSABP).17 The relationship between the RS and the likelihood of chemotherapy response is not yet known.
The goal of the current research was (1) to directly examine the correlation between RS and pCR to preoperative chemotherapy and (2) to identify additional genes that are associated with pCR. We used RT-PCR to quantify the expression of 384 genes, including the RS panel of 21 genes, in FPE diagnostic biopsy specimens of women with locally advanced breast cancer who were treated with chemotherapy before surgery in a study conducted at the Istituto Nazionale Tumori (INT) of Milan (Italy). The genes identified in this cohort of patients were then tested for their correlation with pCR in a separate cohort of 82 patients treated at M.D. Anderson Cancer Center (MDACC; Houston, TX) with similar neoadjuvant chemotherapy whose tumor tissue was profiled by DNA microarrays.
METHODS
INT-Milan Patients
Women with locally advanced breast cancer who enrolled from 1998 to 2002 in a study of primary chemotherapy were eligible. After diagnostic core biopsy and before surgery, patients were treated with doxorubicin (60 mg/m2) and paclitaxel (200 mg/m2) every 3 weeks x 3, followed by weekly paclitaxel (80 mg/m2) x 12. After surgery, adjuvant CMF (cyclophosphamide, methotrexate, and fluorouracil), locoregional irradiation, and hormonal therapy were administered using standard criteria. Core biopsy paraffin blocks were available for 95 of 102 consecutive eligible patients entered onto the trial. Patients were monitored monthly during therapy, and then with physical examination and imaging assessments every 6 months for the first year, and yearly thereafter. This study was performed after approval by the local ethics committee.
Sample Preparation for RT-PCR
Three 10-μm sections from each paraffin block were placed into two bar-coded microcentrifuge tubes (a total of six sections per patient). One additional 5-μm section was stained with hematoxylin and eosin (H&E). Tubes and slides were shipped to Genomic Health Inc (Redwood City, CA) at ambient temperature.
Paraffin was removed by xylene extraction, and RNA was isolated using Epicentre Technologies Inc's (Madison, WI) MasterPure RNA Purification kit. Total RNA was measured using the RiboGreen RNA Quantitation Kit (Molecular Probes Inc). Residual genomic DNA contamination was assayed by TaqMan (Applied Biosystems, Foster City, CA) PCR assay for ?-actin DNA. Samples were treated with additional DNase I if required.
Gene Expression Analysis by RT-PCR
Quantitative gene expression was determined by a multi-analyte TaqMan RT-PCR assay that was designed to accurately measure the small fragments of RNA present in archival tumor blocks.15 Reverse transcription of purified RNA was performed using gene specific primers and random hexamers. TaqMan reactions were performed in 384 well plates, using Applied Biosystems PRISM 7900HT TaqMan instruments.
We selected "candidate" genes by surveying the breast cancer literature for evidence of linkage to either cancer pathological processes (eg, proliferation, invasion, apoptosis, metastasis, angiogenesis, immune surveillance, tumor suppression activity, oncogene activity, and differentiation status) or to pathways important in chemotherapy response (eg, metabolism, drug resistance, transporters, and DNA repair). In addition, we included a number of genes identified in published DNA microarray studies of breast cancer, including genes that mark cellular subtypes,19 that may be prognostically significant,6,20 or that predict response to chemotherapy.21-24
We profiled expression of a total of 384 genes. Expression of each gene was measured in a single well and then normalized relative to the median of all genes. In addition, the expression of 21 genes used in the Breast Cancer Oncotype DX RS assay were measured in triplicate, and five reference genes (ACTB or beta-actin, GAPDH, GUS, RPLPO, and TFRC) were used for normalization.17 Normalized expression measurements typically range from 0 to 15, where a one-unit change generally reflects a two-fold change in RNA.
Pathology Assessments
The primary end point of pCR was determined by pathology assessment of the surgical specimen. We defined pCR as the absence of invasive cancer in the breast (residual DCIS was allowed). H&E-stained slides submitted to Genomic Health were reviewed by a pathologist for confirmation of the diagnosis, and for assessment of the proportion of tissue surface area composed of tumor. Immunohistochemistry (IHC) for estrogen receptor (ER) and progesterone receptor (PR) was performed using reagents from Lab Vision-Neomarkers (Fremont, CA). The threshold for categorizing IHC as positive was 10% of cells. Tumor grade was assessed using the Bloom-Richardson criteria.25
Statistical Methods for RT-PCR Analyses
Demographic and baseline characteristics were summarized by descriptive statistics. Probit regression of gene expression and pCR was performed in a univariate analysis for all 384 tested candidate genes. Because the number of candidate genes relative to the number of patients is large, we performed simulations in which we assigned pCR to 11 randomly selected patients to estimate the number of genes that would appear to be significant in the absence of a genuine association with pCR (false-positive rate). Cluster analysis was performed on the genes that were significantly associated with pCR in the univariate analysis in order to identify gene groups that were related to chemotherapy response. Forward stepwise regression of pCR as a function of gene expression was performed, and bootstrap resampling was performed to assess the robustness of the stepwise procedure.
MDACC-Houston Patients and DNA Microarray Analysis
DNA microarray results from 82 patients treated at the Nellie B. Connally Breast Center of The University of Texas MDACC were used to assess the response discriminating value of the pCR-associated genes in independent data. Gene expression profiling was performed on fine-needle aspirations (FNA) of breast cancers that were collected prospectively for an ongoing pharmacogenomic marker discovery study. FNA was performed using 23- or 25-gauge needle before starting preoperative chemotherapy with 12 weeks of paclitaxel followed by fluorouracil, doxorubicin and cyclophosphamide (FAC) x 4 courses. Cells from two to three passes were collected into vials containing 1 mL of RNAlater solution (Ambion, Austin, TX) and stored at –80°C. At the completion of neoadjuvant chemotherapy all patients had surgical resection of the tumor bed, and all areas of radiologically and/or architecturally abnormal tissue were submitted for histopathologic examination. Pathologic CR was defined as no residual invasive cancer in the breast (residual DCIS was allowed) and lymph nodes. This study was approved by MDACC's institutional review board (IRB) and all patients signed an informed consent for voluntary participation. At the time of this analysis, the first 85 cases of the study were hybridized, three chips failed the quality check (QC) process, and therefore gene expression profiles from 82 samples were available for this collaboration.
RNA was extracted from FNA samples using the RNAeasy Kit (Qiagen, Valencia, CA). The amount and quality of RNA was determined with a DU-640 UV Spectrophotometer (Beckman Coulter, Fullerton, CA) and it was considered adequate for further analysis if the OD 260/280 ratio was 1.8 and the total RNA yield was 1 μg. Median RNA yield of the 85 specimens was 2.0 μg with a range of 1 to 22 μg. The cellular composition of the FNA samples was reported previously.26 In brief, FNA samples on average contain 80% neoplastic cells and the rest of the cells are infiltrating leukocytes. These samples contain little or no stromal cells (eg, fibroblasts or adipocytes) or normal breast epithelium. Complementary RNA (cRNA) was generated using a standard Affymetrix protocol. A single round of RNA amplification was carried out. For each patient, the fragmented cRNA was hybridized to an Affymetrix U133A GeneChip array (Santa Clara, CA), overnight at 42°C. This chip contains 22,215 different probe sets that correspond to 13,739 human UniGene clusters (genes). Hybridization cocktail was prepared as described in the Affymetrix technical manual. dCHIP V1.3 (http://dchip.com) software was used to generate probe level intensities and quality measures, including median intensity, percentage of probe set outliers and percentage of single probe outliers for each chip. Three chips failed the QC process and subsequent analysis was performed on 82 samples.
The dCHIP software was used for normalization; this program normalizes all arrays to one standard array that represents a chip with median overall intensity. After normalization, probe set level intensity estimates were generated as follows. Estimates of feature level intensity was derived from the 75th percentile of each features' pixel level intensities. Each individual probe is aggregated at the feature level to form a single measure of intensity for each probe set. We used the perfect match model. Normalized gene expression values were transformed to the log-scale (base 10) for analysis.
To identify informative genes differentially expressed between cases with pCR and those with residual disease, genes were ordered by P values obtained with two-sample, unequal-variance t tests. As a conservative approach for calling statistical association of gene expression and pCR, it was required that each Affymetrix probe present for a single gene demonstrate statistical significance. To assess the response discriminating value of the genes identified from the RT-PCR data, hierarchical clustering and multidimensional scaling of the microarray data was performed. Central linkage hierarchical clustering was performed with 1-Pearson correlation coefficient as distance metric. Cluster reproducibility and the robustness of the dendograms were examined by the method proposed by McShane et al based on 1,000 perturbations.18 The correlation between robust molecular clusters and pCR as well as classic histopathologic characteristics were examined by the 2 test. Multidimensional scaling (MDS) was performed to determine whether the expression profiles are clustered (rather than represent the same multivariate Gaussian distribution) and whether the clusters correlate with response, with the Euclidian distance as metric. Statistical analysis was performed by using the BRB-Arraytools software package (http://linus.nci.nih.gov/BRB-ArrayTools.html).
RESULTS
Sample Disposition and Patient Characteristics at INT-Milan
Ninety-five core biopsy tumor specimens were submitted for gene expression analysis by RT-PCR. Pathology review indicated that two samples had little or no tumor tissue (< 5% of the section) and were not assessable. The yield of total RNA was insufficient to assay (< 500 ng) in four patients. RT-PCR profiles were within the specifications of the assay for all patients.
Characteristics at diagnosis for the 89 assessable patients in Milan are listed in Table 1. Seventy-six patients had infiltrating ductal carcinoma. The mean tumor size at diagnosis was 6.4 cm (standard deviation, 2.3 cm).
The outcomes following primary chemotherapy are summarized in Table 2. Overall, 11 patients (12%) had a pCR.
Patients whose tumors were ER-negative by IHC analysis of the core biopsy specimen appeared to be more likely to have a pCR. Seven (23%) of the 31 patients (95% CI, 8% to 37%) with ER-negative breast cancer by IHC had a pCR. Four (8%) of the 52 patients (95% CI, 1% to 15%) with ER-positive breast cancer by IHC had a pCR. There was no correlation between patient age or tumor grade and pCR.
IHC and RT-PCR for ER and PR in INT-Milan Patients
IHC and RT-PCR measurements of tumor expression of ER and PR were compared. The concordance between RT-PCR (mRNA) and IHC (protein) was high for both ER ( = .84; 95% CI, 0.71 to 0.96) and PR ( = .71; 95% CI, 0.56 to 0.86).
Univariate Analysis of Gene Expression and pCR in INT-Milan Patients
The odds ratio for pCR for each of the candidate genes with P < .05 are given in Table 3. Thirty genes had a P value of < .01 and a total of 86 genes had a P value < .05. Simulations based on randomly assigning pCR to 11 patients indicated that four genes would be expected to have a P value of < .01 and 19 genes have a P value < .05 by chance alone.
A number of the genes that correlated with pCR fell into categories defined by biologic function, prominently: proliferation (eg, MCM6, E2F1, MYBL2), apoptosis (BBC3, BAD, DR4, TP53BP1), invasion and metastasis (FYN and MMP12), and drug resistance/metabolism (ABCC5, ALDH1A1, CYP3A4).
The quantitative expression of the ER gene by RT-PCR significantly correlated with the likelihood of pCR with an odds ratio of 0.78 (a 22% decrease in the likelihood of pCR with each increment of one unit in ER expression).
Correlated Gene Expression in INT-Milan Patients
Unsupervised cluster analysis of the 86 genes that were associated with pCR in univariate analysis was performed to identify groups of co-expressed genes that are related to the likelihood of pCR (Fig 1). Three prominent groups of such co-expressed genes were identified—an ER-related gene cluster, a proliferation-related gene cluster, and an immune-related gene cluster. Higher expression of an ER gene cluster (including PR, SCUBE2, ER, NPD009, GATA3, IGF1R, IRS1) correlated with a lower likelihood of pCR. On the other hand, higher expression of a proliferation gene cluster (including CDC20, E2F1, MYBL2, TOP2A, FBXO5, MCM2, MCM6, CDC25B) or of an immune-related gene cluster (including MCP1, CD68, CTSB, CD18, ILT-2, CD3z, FasL, HLA.DPB1, GBP1) correlated with a higher likelihood of pCR.
The fact that a substantial percentage (22%) of tested candidate genes correlated with pCR in univariate analysis is partially explained by their co-expression. Forty-six of the 86 "hits" with respect to pCR were in either the ER cluster (defined by r > 0.45, using GATA3 as reference), the proliferation cluster (defined by r > 0.45 using MCM6 as reference) or the immune response cluster (defined by r > 0.45 using ILT-2 as a reference).
On the other hand, a number of genes that did not belong to one of these three clusters of co-expressed genes also correlated with pCR. For example, higher expression of the ?-tubulin gene (TUBB), the target of paclitaxel, was significantly correlated with a lower likelihood of pCR, as was expression of the DNA repair genes MSH3 and ERCC1. Higher expression of the transcriptional regulator ID2 was correlated with higher likelihood of pCR (odds ratio, 2.37; P = .0026).
Multivariate Analysis of Gene Expression and pCR in INT-Milan Patients
Multivariate analysis was performed to identify multi-gene models for future testing. Due to the limited number of pathologically complete responders under study (11 pCRs) and the large number of genes being tested (384 genes), many multiple-gene models can discriminate pCR. For instance, a three-gene model with TBP, ILT-2 and MCM6 generated predicted probabilities of pCR that were nearly 100% concordant with observed probabilities. Consequently, bootstrap resampling,27 based on 1,000 resamplings with replacement of the original dataset, was performed to assess the arbitrariness of stepwise variable selection procedures based on the probit model. Using a 0.1 criterion for both entry and exit Wald 2 probabilities for an effect, results revealed that 26% of the selected models included only a single gene with an average concordance between predicted and observed probability of pCR greater than 90%. Furthermore, all multi-gene models contained no more than four genes and yielded an average concordance between predicted and observed probability of pCR of 97%. Consequently, multigene models should be considered exploratory and interpreted with caution.
RS and pCR in INT-Milan Patients
RS is calculated from the expression of 21 genes (16 cancer-related genes and five reference genes) as measured by the RT-PCR.17 The RS (on a scale from 0 to 100) is a function of the expression of four groups of genes plus three individual genes: an ER-related group—ER, PR, Bcl2, and SCUBE2; a proliferation group—Ki-67, MYBL2, Survivin, Cyclin B1, and STK15; a HER2 group—HER2 and GRB7; an invasion group—Stromelysin 3 and Cathepsin L2; and the three individual genes—GSTM1, BAG1, and CD68.
Probit regression,28 based on iteratively reweighted least squares, was used to assess the correlation of RS to pCR. A global likelihood ratio test comparing the (full) model with RS to the (null) model excluding RS was performed, resulting in a P value of .005. Model assessments, including assessments of goodness of fit27 (Hosmer and Lemeshow test, P = .83) and residual analysis did not reveal significant model departures.
Additionally, model validation29 was performed to confirm that the results were not spurious and to assess the performance of the RS model with respect to discrimination and potential overfitting. Bootstrap resampling,30 based on 1,000 resamplings with replacement of the original data set, yielded bias-corrected measures of model discrimination31-33 (Somer's D rank correlation, 0.50; generalized R2 = 0.14; calibrated slope, 1.13) indicating that the RS provides good discrimination of pCR. Empirical 0.5% and 99.5% percentiles for the RS regression coefficient were 0.001 and 0.045, respectively, further suggesting that the model is predictive of pCR. Predicted values of the probability of pCR as a function of RS derived from the probit model fit and normal two-sided 95% CIs are provided in Figure 2. The probability of pCR increased with RS.
Assessment of the 86 Informative Genes in a Second Independent Patient Cohort
Gene expression data from 82 patients with stage I-III breast cancer was used to assess the response discriminating value of the 86 genes identified by RT-PCR. All of these patients received sequential paclitaxel followed by FAC preoperative chemotherapy at MDACC. Characteristics at diagnosis for the 82 assessable patients treated at MDACC in Houston and their outcomes after neoadjuvant chemotherapy are given in Tables 1 and 2, respectively. Overall, 21 patients (26%) had a pCR.
Seventy-nine of the 86 genes identified as predictive in the INT-Milan cohort were represented on the Affymetrix U133A chip by 179 individual probe sets. The seven genes not on the array were AK055699, CCND1, Contig 51037, DR4, IGFBP5, KIAA1209, and RHOC.
The correlation of each probe set with pCR is shown in Table 4. Thirty-seven probe sets corresponding to 24 genes demonstrated correlation with pCR with P .05 (expected false discovery, four genes). Of the 79 genes, 32 (40%) showed association with pCR at P .1 (expected false discovery, eight genes). In all cases, the association of expression with pCR was in the same direction in the MDACC patients as had been observed in the INT-Milan patients.
We also performed supervised hierarchical clustering using all 179 probe sets to see if the cases fall into distinct molecular clusters that correspond to different rates of pCR. The resulting dendogram had two main clusters that were robust and reproducible (R-index = 0.938, D-index = 1.675; Fig 3). One cluster group with 39 cases included five pCRs (pCR rate, 12%), the other cluster had 43 cases including 16 pCRs (37%), P = .01. The cluster groups also correlated with clinical ER status assessed by immunohistochemistry (P < .001). Multidimensional scaling with all 179-probe sets also revealed that cases with pCR clustered together and the Global Test for significance yielded P < .001. These results indicate that many of the genes that were associated with pCR in the RT-PCR data from Milan had response-discriminating value in an independent data set even when measured with a different gene expression profiling platform.
DISCUSSION
We have identified a set of genes that in their expression correlate with pCR to neoadjuvant doxorubicin and paclitaxel. Of 384 candidate genes tested by RT-PCR analysis, the expression of 86 significantly correlated with pCR (P < .05). Pathologic complete response is less frequent than major clinical response, but is the independent variable most strongly associated with the likelihood of long-term disease-free survival.1,2,4
Because this study tested hundreds of candidate genes, a number of the 86 predictive genes may be false discoveries. Nevertheless, it is highly likely that many of them are valid. First, simulations that assign pCR to 11 patients at random show that at P < .05 it is expected to find only 19 positive genes by chance alone. The findings are also biologically plausible since these genes represent functional categories that are reported to influence sensitivity and resistance to chemotherapy (ie, apoptosis, invasion, metastasis, drug resistance/metabolism, proliferation, ER). Importantly, the response discriminating value of these genes was confirmed in an independent data set examined with a different gene expression profiling technology. Twenty-four (30%) of the 86 genes showed similar correlation with pCR (P < .05) in the microarray data. An additional eight genes showed modest correlation with pCR (P < .1). All of these genes had the same directional association with pCR in both cohorts of patients. Thus, many genes identified by RT-PCR had response discriminating value when measured by oligonucleotide arrays. This is notable because the two platforms have very different sensitivity and dynamic range to measure mRNA expression.34,35 Also, the tissues used for discovery in Milan were formalin-fixed, paraffin-embedded core-needle biopsies, whereas the tissues used in the confirmation analysis were obtained with FNA. These technical differences could partly explain why fewer than half of the genes showed high correlation with pCR in both data sets. The fact that a number of genes retained their predictive value in totally different groups of patients undergoing similar chemotherapy in different cancer centers seems to fulfill the type of validation recently advocated to minimize the risk of "noise" discovery.36
The 86 predictive genes included three particularly prominent co-expression clusters. Pathologic complete response correlated with higher expression of genes regulating proliferation (eg, CDC20, E2F1, MYBL2, TOPO2A) and immune response (eg, MCP1, CD68, CTSB, CD18, ILT-2, CD3z, FasL, HLA.DPB1), and with lower expression of genes related to the expression of the ER (eg, ER, PR, SCUBE2, and GATA3). These relationships of proliferation and ER to chemotherapy response are consistent with several previously reported observations.37-40
The observation that higher expression of an immune-related gene cluster is associated with chemotherapy benefit is remarkable. It is possible that the pretreatment host response may enhance the ability of chemotherapy to eliminate cancer cells. This observation should be evaluated in future studies.
It is noteworthy that the multi-drug resistance gene MDR-1, which has been frequently implicated in chemotherapy resistance in vitro,41-43 was not related to chemotherapy response here. MDR-1 also did not link to clinical anthracycline resistance in a recent paper by Faneyte et al.44 One of the genes (CYBA) previously identified by Chang et al was correlated with pCR in this study.21
Genes related to proliferation (Ki-67, Survivin, MYBL2, STK15, Cyclin B1) and the ER (ER, PGR, Bcl2, SCUBE2) are significantly represented in the 21 gene panel employed by the gene expression–based RS assay.17 The RS has been validated to quantify the risk of recurrence in tamoxifen-treated patients with node-negative, ER-positive breast cancer. We show here that RS strongly correlated with pCR. This has a provocative clinical implication, namely, that patients with high RS values, who are most likely to experience recurrence,17 are the very patients most likely to receive the greatest clinical benefit from chemotherapy treatment. It will be important to further evaluate this concept in future studies.
In conclusion, this study correlated the expression of certain genes to the probability of pathologic complete response. The observation that many of the discriminating genes maintained their predictive value across two cohorts and two different gene expression profiling platforms makes the findings compelling enough to subject these genes to further clinical evaluation. The results of this study are consistent with a growing body of data that indicates that not all breast cancer patients respond equally to chemotherapy. Successful prediction of a durable benefit of chemotherapy using routinely obtained fixed tumor tissue would make possible and widely applicable the tailoring of toxic chemotherapy to individual patients.
Authors' Disclosures of Potential Conflicts of Interest
Although all authors completed the disclosure declaration, the following authors or their immediate family members indicated a financial interest. No conflict exists for drugs or devices used in a study if they are not being evaluated as part of the investigation. For a detailed description of the disclosure categories, or for more information about ASCO's conflict of interest policy, please refer to the Author Disclosure Declaration and the Disclosures of Potential Conflicts of Interest section in Information for Contributors.
Acknowledgment
We acknowledge John Morlan for his assistance with RNA extraction, Chithra Sangli for biostatistical input, and Michael Kiefer for help with candidate gene selection.
NOTES
Supported in part by a grant from Genomic Health Inc, Redwood City, CA.
Presented in abstract form at the 40th Annual Meeting of American Society of Clinical Oncology, New Orleans, LA, June 5-8, 2004.
Terms in blue are defined in the glossary, found at the end of this issue and online at www.jco.org.
Authors' disclosures of potential conflicts of interest are found at the end of this article.
REFERENCES
Fisher B, Bryant J, Wolmark N, et al: Effect of preoperative chemotherapy on the outcome of women with operable breast cancer. J Clin Oncol 16:2672-2685, 1998
Bonadonna G, Valagussa P, Brambilla C, et al: Primary chemotherapy in operable breast cancer: Eight-year experience at the Milan Cancer Institute. J Clin Oncol 16:93-100, 1998
Smith IC, Heys SD, Hutcheon AW, et al: Neoadjuvant chemotherapy in breast cancer: Significantly enhanced response with docetaxel. J Clin Oncol 20:1456-1466, 2002
Kaufmann M, von Minckwitz G, Smith R, et al: International expert panel on the use of primary (preoperative) systemic treatment of operable breast cancer: Review and recommendations. J Clin Oncol 21:2600-2608, 2003
Rouzier R, Wagner P, Morandi P, et al: Gene expression profiling of primary breast cancer. Curr Oncol Rep 7:38-44, 2005
Sorlie T, Perou CM, Tibshirani R, et al: Gene expression patterns of breast carcinomas distinguish tumor subclasses with clinical implications. Proc Natl Acad Sci USA 98:10869-10874, 2001
van de Vijver MJ, He YD, van't Veer LJ, et al: A gene-expression signature as a predictor of survival in breast cancer. N Engl J Med 347:1999-2009, 2002
Chang J, Powles TJ, Allred DC, et al: Biologic markers as predictors of clinical outcome from systemic therapy for primary operable breast cancer. J Clin Oncol 17:3058-3063, 1999
Sotiriou C, Powles TJ, Dowsett M, et al: Gene expression profiles derived from fine needle aspiration correlate with response to systemic chemotherapy in breast cancer. Breast Cancer Res 4:R3, 2002
Assersohn L, Gangi L, Zhao Y, et al: The feasibility of using fine needle aspiration from primary breast cancers for cDNA microarray analyses. Clin Cancer Res 8:794-801, 2002
Ayers M, Symmans WF, Stec J, et al: Gene expression profiles predict complete pathologic response to neoadjuvant paclitaxel and fluorouracil, doxorubicin, and cyclophosphamide chemotherapy in breast cancer. J Clin Oncol 22:2284-2293, 2004
Ellis M, Davis N, Coop A, et al: Development and validation of a method for using breast core needle biopsies for gene expression microarray analyses. Clin Cancer Res 8:1155-1166, 2002
Karsten SL, Van Deerlin VM, Sabatti C, et al: An evaluation of tyramide signal amplification and archived fixed and frozen tissue in microarray gene expression analysis. Nucleic Acids Res 30:E4, 2002
Ramaswamy S, Golub TR: DNA microarrays in clinical oncology. J Clin Oncol 20:1932-1941, 2002
Cronin M, Pho M, Dutta D, et al: Measurement of gene expression in archival paraffin-embedded tissues: Development and performance of a 92-gene reverse transcriptase-polymerase chain reaction assay. Am J Pathol 164:35-42, 2004
Specht K, Richter T, Muller U, et al: Quantitative gene expression analysis in microdissected archival formalin-fixed and paraffin-embedded tumor tissue. Am J Pathol 158:419-429, 2001
Paik S, Shak S, Tang G, et al: A multigene assay to predict recurrence of tamoxifen-treated, node-negative breast cancer. N Engl J Med 351:2817-2826, 2004
McShane LM, Radmacher MD, Simon R: Methods for assessing reproducibility of clustering patterns observed in analyses of microarray data. Bioinformatics 18:1462-1469, 2002
Perou CM, Sorlie T, Eisen MB, et al: Molecular portraits of human breast tumours. Nature 406:747-752, 2000
van't Veer LJ, Dai H, van de Vijver MJ, et al: Gene expression profiling predicts clinical outcome of breast cancer. Nature 2002;415(6871):530-6.
Chang JC, Wooten EC, Tsimelzon A, et al: Gene expression profiling for the prediction of therapeutic response to docetaxel in patients with breast cancer. Lancet 362:362-369, 2003
Duran GE, Wang YC, Francisco B, et al: Identification of genes associated with non-MDR1 taxane resistance in human ovarian carcinoma cell lines by microarray analysis. Am Assoc Cancer Res, 2003 (abstr 5760)
Lamendola DE, Duan Z, Yusuf RZ, et al: Molecular description of evolving paclitaxel resistance in the SKOV-3 human ovarian carcinoma cell line. Cancer Res 63:2200-2205, 2003
Duan Z, Feller AJ, Penson RT, et al: Discovery of differentially expressed genes associated with paclitaxel resistance using cDNA array technology: analysis of interleukin (IL) 6, IL-8, and monocyte chemotactic protein 1 in the paclitaxel-resistant phenotype. Clin Cancer Res 5:3445-3453, 1999
Elston CW, Ellis IO: Pathological prognostic factors in breast cancer, I: The value of histological grade in breast cancer: experience from a large study with long-term follow-up. Histopathology 19:403-410, 1991
Symmans WF, Ayers M, Clark EA, et al: Total RNA yield and microarray gene expression profiles from fine-needle aspiration biopsy and core-needle biopsy samples of breast carcinoma. Cancer 97:2960-2971, 2003
Hosmer DW, Lemeshow S: Applied Logistic Regression. New York, NY, Wiley, 1989
B. J. T. Morgan: Analysis of Quantal Response Data. New York, NY, Chapman and Hall, 1992
Harrel FE Jr: Regression Modeling Strategies: With Applications to Linear Models, Logistic Regression, and Survival Analysis. New York, NY, Springer-Verlag, 2001
Efron B, Tibshirani R: An Introduction to the Bootstrap. New York, NY, Chapman and Hall, 1993
Somers RH: A new asymmetric measure of association for ordinal variables. Am Sociol Rev 27:799-811, 1962
Nagelkerke NJD: A note on a general definition of the coefficient of determination. Biometrika 78:691-692, 1991
Efron B: Estimating the error rate of a prediction rule: Improvement on cross-validation. J Am Stat Assoc 78:316-331, 1983
Kuo WP, Jenssen TK, Butte AJ, et al: Analysis of matched mRNA measurements from two different microarray technologies. Bioinformatics 18:405-412, 2002
Yuen T, Wurmbach E, Pfeffer RL, et al: Accuracy and calibration of commercial oligonucleotide and custom cDNA microarrays. Nucleic Acids Res 30:e48, 2002
Ioannidis JP: Microarrays and molecular research: Noise discovery? Lancet 365:454-455, 2005
Lippman ME, Allegra JC, Thompson EB, et al: The relation between estrogen receptors and response rate to cytotoxic chemotherapy in metastatic breast cancer. N Engl J Med 298:1223-1228, 1978
Kuerer HM, Newman LA, Smith TL, et al: Clinical course of breast cancer patients with complete pathologic primary tumor and axillary lymph node response to doxorubicin-based neoadjuvant chemotherapy. J Clin Oncol 17:460-469, 1999
Colleoni M, Minchella I, Mazzarol G, et al: Response to primary chemotherapy in breast cancer patients with tumors not expressing estrogen and progesterone receptors. Ann Oncol 11:1057-1059, 2000
Gianni L, Baselga J, Eiermann W, et al: First report of the European Cooperative Trial in operable breast cancer (ECTO): Effects of primary systemic therapy (PST) on local-regional disease. Proc Am Soc Clin Oncol, 2002 (abstr 132)
Penson RT, Oliva E, Skates SJ, et al: Expression of multidrug resistance-1 protein inversely correlates with paclitaxel response and survival in ovarian cancer patients: A study in serial samples. Gynecol Oncol 93:98-106, 2004
Wu H, Hait WN, Yang JM: Small interfering RNA-induced suppression of MDR1 (P-glycoprotein) restores sensitivity to multidrug-resistant cancer cells. Cancer Res 63:1515-1519, 2003
Burger H, Foekens JA, Look MP, et al: RNA expression of breast cancer resistance protein, lung resistance-related protein, multidrug resistance-associated proteins 1 and 2, and multidrug resistance gene 1 in breast cancer: Correlation with chemotherapeutic response. Clin Cancer Res 9:827-836, 2003
Faneyte IF, Kristel PM, van de Vijver MJ: Multidrug resistance associated genes MRP1, MRP2 and MRP3 in primary and anthracycline exposed breast cancer. Anticancer Res 24:2931-2939, 2004(Luca Gianni, Milvia Zambe)
The University of Texas M.D. Anderson Cancer Center, Houston, TX: Genomic Health Inc, Redwood City, CA
Millennium Pharmaceuticals Inc, Cambridge, MA
ABSTRACT
PURPOSE: We sought to identify gene expression markers that predict the likelihood of chemotherapy response. We also tested whether chemotherapy response is correlated with the 21-gene Recurrence Score assay that quantifies recurrence risk.
METHODS: Patients with locally advanced breast cancer received neoadjuvant paclitaxel and doxorubicin. RNA was extracted from the pretreatment formalin-fixed paraffin-embedded core biopsies. The expression of 384 genes was quantified using reverse transcriptase polymerase chain reaction and correlated with pathologic complete response (pCR). The performance of genes predicting for pCR was tested in patients from an independent neoadjuvant study where gene expression was obtained using DNA microarrays.
RESULTS: Of 89 assessable patients (mean age, 49.9 years; mean tumor size, 6.4 cm), 11 (12%) had a pCR. Eighty-six genes correlated with pCR (unadjusted P < .05); pCR was more likely with higher expression of proliferation-related genes and immune-related genes, and with lower expression of estrogen receptor (ER) –related genes. In 82 independent patients treated with neoadjuvant paclitaxel and doxorubicin, DNA microarray data were available for 79 of the 86 genes. In univariate analysis, 24 genes correlated with pCR with P < .05 (false discovery, four genes) and 32 genes showed correlation with P < .1 (false discovery, eight genes). The Recurrence Score was positively associated with the likelihood of pCR (P = .005), suggesting that the patients who are at greatest recurrence risk are more likely to have chemotherapy benefit.
CONCLUSION: Quantitative expression of ER-related genes, proliferation genes, and immune-related genes are strong predictors of pCR in women with locally advanced breast cancer receiving neoadjuvant anthracyclines and paclitaxel.
INTRODUCTION
When chemotherapy is administered preoperatively to women with breast cancer, a rate of clinical complete and partial response exceeding 70% is achieved with many regimens.1-3 However, a lower percentage ranging from about 5% to 38% of patients achieve tumor eradication when response is assessed with careful pathologic examination of the surgical specimen after chemotherapy (pathologic complete response [pCR]).1-4 Several studies show that, together with nodal status at surgery, pCR stands out as the most significant independent variable associated with the likelihood of benefit as measured by disease-free and overall survival.4 In principle, diagnostic tests that predict which patients will experience a pCR will identify patients who will likely benefit from treatment, and, will also indentify individuals with low or no likelihood of benefit, sparing them the toxicity of an inactive treatment and allowing earlier application of alternative approaches.
Over the last several years a number of laboratories have demonstrated that DNA microarray–based gene expression profiling (GEP) of patient tumor tissue provides prognostic and predictive information.5 In breast cancer in particular, several DNA microarray studies suggest that GEP can provide better prognostic discrimination than traditional clinical tools, and moreover can predict response to preoperative chemotherapy.6-12 Continued progress with DNA microarrays in this area, toward biomarker discovery, clinical validation, and ultimately routine clinical use, is limited by DNA microarray requirements for fresh tumor tissue.13,14
Recent work shows that real time reverse transcriptase polymerase chain reaction (RT-PCR) can quantify gene expression in formalin-fixed paraffin-embedded (FPE) tumor tissue specimens with high reproducibility, specificity and sensitivity.15,16 Using this technique to measure expression of 250 candidate genes, three studies in a total of 447 patients were carried out to identify a 21-gene panel and Recurrence Score (RS) algorithm. The RS was clinically validated to quantify the risk of distant recurrence in 668 women with estrogen receptor–positive, axillary lymph node-negative breast tumors treated with adjuvant tamoxifen in a clinical trial conducted by the National Surgical Adjuvant Breast and Bowel Project (NSABP).17 The relationship between the RS and the likelihood of chemotherapy response is not yet known.
The goal of the current research was (1) to directly examine the correlation between RS and pCR to preoperative chemotherapy and (2) to identify additional genes that are associated with pCR. We used RT-PCR to quantify the expression of 384 genes, including the RS panel of 21 genes, in FPE diagnostic biopsy specimens of women with locally advanced breast cancer who were treated with chemotherapy before surgery in a study conducted at the Istituto Nazionale Tumori (INT) of Milan (Italy). The genes identified in this cohort of patients were then tested for their correlation with pCR in a separate cohort of 82 patients treated at M.D. Anderson Cancer Center (MDACC; Houston, TX) with similar neoadjuvant chemotherapy whose tumor tissue was profiled by DNA microarrays.
METHODS
INT-Milan Patients
Women with locally advanced breast cancer who enrolled from 1998 to 2002 in a study of primary chemotherapy were eligible. After diagnostic core biopsy and before surgery, patients were treated with doxorubicin (60 mg/m2) and paclitaxel (200 mg/m2) every 3 weeks x 3, followed by weekly paclitaxel (80 mg/m2) x 12. After surgery, adjuvant CMF (cyclophosphamide, methotrexate, and fluorouracil), locoregional irradiation, and hormonal therapy were administered using standard criteria. Core biopsy paraffin blocks were available for 95 of 102 consecutive eligible patients entered onto the trial. Patients were monitored monthly during therapy, and then with physical examination and imaging assessments every 6 months for the first year, and yearly thereafter. This study was performed after approval by the local ethics committee.
Sample Preparation for RT-PCR
Three 10-μm sections from each paraffin block were placed into two bar-coded microcentrifuge tubes (a total of six sections per patient). One additional 5-μm section was stained with hematoxylin and eosin (H&E). Tubes and slides were shipped to Genomic Health Inc (Redwood City, CA) at ambient temperature.
Paraffin was removed by xylene extraction, and RNA was isolated using Epicentre Technologies Inc's (Madison, WI) MasterPure RNA Purification kit. Total RNA was measured using the RiboGreen RNA Quantitation Kit (Molecular Probes Inc). Residual genomic DNA contamination was assayed by TaqMan (Applied Biosystems, Foster City, CA) PCR assay for ?-actin DNA. Samples were treated with additional DNase I if required.
Gene Expression Analysis by RT-PCR
Quantitative gene expression was determined by a multi-analyte TaqMan RT-PCR assay that was designed to accurately measure the small fragments of RNA present in archival tumor blocks.15 Reverse transcription of purified RNA was performed using gene specific primers and random hexamers. TaqMan reactions were performed in 384 well plates, using Applied Biosystems PRISM 7900HT TaqMan instruments.
We selected "candidate" genes by surveying the breast cancer literature for evidence of linkage to either cancer pathological processes (eg, proliferation, invasion, apoptosis, metastasis, angiogenesis, immune surveillance, tumor suppression activity, oncogene activity, and differentiation status) or to pathways important in chemotherapy response (eg, metabolism, drug resistance, transporters, and DNA repair). In addition, we included a number of genes identified in published DNA microarray studies of breast cancer, including genes that mark cellular subtypes,19 that may be prognostically significant,6,20 or that predict response to chemotherapy.21-24
We profiled expression of a total of 384 genes. Expression of each gene was measured in a single well and then normalized relative to the median of all genes. In addition, the expression of 21 genes used in the Breast Cancer Oncotype DX RS assay were measured in triplicate, and five reference genes (ACTB or beta-actin, GAPDH, GUS, RPLPO, and TFRC) were used for normalization.17 Normalized expression measurements typically range from 0 to 15, where a one-unit change generally reflects a two-fold change in RNA.
Pathology Assessments
The primary end point of pCR was determined by pathology assessment of the surgical specimen. We defined pCR as the absence of invasive cancer in the breast (residual DCIS was allowed). H&E-stained slides submitted to Genomic Health were reviewed by a pathologist for confirmation of the diagnosis, and for assessment of the proportion of tissue surface area composed of tumor. Immunohistochemistry (IHC) for estrogen receptor (ER) and progesterone receptor (PR) was performed using reagents from Lab Vision-Neomarkers (Fremont, CA). The threshold for categorizing IHC as positive was 10% of cells. Tumor grade was assessed using the Bloom-Richardson criteria.25
Statistical Methods for RT-PCR Analyses
Demographic and baseline characteristics were summarized by descriptive statistics. Probit regression of gene expression and pCR was performed in a univariate analysis for all 384 tested candidate genes. Because the number of candidate genes relative to the number of patients is large, we performed simulations in which we assigned pCR to 11 randomly selected patients to estimate the number of genes that would appear to be significant in the absence of a genuine association with pCR (false-positive rate). Cluster analysis was performed on the genes that were significantly associated with pCR in the univariate analysis in order to identify gene groups that were related to chemotherapy response. Forward stepwise regression of pCR as a function of gene expression was performed, and bootstrap resampling was performed to assess the robustness of the stepwise procedure.
MDACC-Houston Patients and DNA Microarray Analysis
DNA microarray results from 82 patients treated at the Nellie B. Connally Breast Center of The University of Texas MDACC were used to assess the response discriminating value of the pCR-associated genes in independent data. Gene expression profiling was performed on fine-needle aspirations (FNA) of breast cancers that were collected prospectively for an ongoing pharmacogenomic marker discovery study. FNA was performed using 23- or 25-gauge needle before starting preoperative chemotherapy with 12 weeks of paclitaxel followed by fluorouracil, doxorubicin and cyclophosphamide (FAC) x 4 courses. Cells from two to three passes were collected into vials containing 1 mL of RNAlater solution (Ambion, Austin, TX) and stored at –80°C. At the completion of neoadjuvant chemotherapy all patients had surgical resection of the tumor bed, and all areas of radiologically and/or architecturally abnormal tissue were submitted for histopathologic examination. Pathologic CR was defined as no residual invasive cancer in the breast (residual DCIS was allowed) and lymph nodes. This study was approved by MDACC's institutional review board (IRB) and all patients signed an informed consent for voluntary participation. At the time of this analysis, the first 85 cases of the study were hybridized, three chips failed the quality check (QC) process, and therefore gene expression profiles from 82 samples were available for this collaboration.
RNA was extracted from FNA samples using the RNAeasy Kit (Qiagen, Valencia, CA). The amount and quality of RNA was determined with a DU-640 UV Spectrophotometer (Beckman Coulter, Fullerton, CA) and it was considered adequate for further analysis if the OD 260/280 ratio was 1.8 and the total RNA yield was 1 μg. Median RNA yield of the 85 specimens was 2.0 μg with a range of 1 to 22 μg. The cellular composition of the FNA samples was reported previously.26 In brief, FNA samples on average contain 80% neoplastic cells and the rest of the cells are infiltrating leukocytes. These samples contain little or no stromal cells (eg, fibroblasts or adipocytes) or normal breast epithelium. Complementary RNA (cRNA) was generated using a standard Affymetrix protocol. A single round of RNA amplification was carried out. For each patient, the fragmented cRNA was hybridized to an Affymetrix U133A GeneChip array (Santa Clara, CA), overnight at 42°C. This chip contains 22,215 different probe sets that correspond to 13,739 human UniGene clusters (genes). Hybridization cocktail was prepared as described in the Affymetrix technical manual. dCHIP V1.3 (http://dchip.com) software was used to generate probe level intensities and quality measures, including median intensity, percentage of probe set outliers and percentage of single probe outliers for each chip. Three chips failed the QC process and subsequent analysis was performed on 82 samples.
The dCHIP software was used for normalization; this program normalizes all arrays to one standard array that represents a chip with median overall intensity. After normalization, probe set level intensity estimates were generated as follows. Estimates of feature level intensity was derived from the 75th percentile of each features' pixel level intensities. Each individual probe is aggregated at the feature level to form a single measure of intensity for each probe set. We used the perfect match model. Normalized gene expression values were transformed to the log-scale (base 10) for analysis.
To identify informative genes differentially expressed between cases with pCR and those with residual disease, genes were ordered by P values obtained with two-sample, unequal-variance t tests. As a conservative approach for calling statistical association of gene expression and pCR, it was required that each Affymetrix probe present for a single gene demonstrate statistical significance. To assess the response discriminating value of the genes identified from the RT-PCR data, hierarchical clustering and multidimensional scaling of the microarray data was performed. Central linkage hierarchical clustering was performed with 1-Pearson correlation coefficient as distance metric. Cluster reproducibility and the robustness of the dendograms were examined by the method proposed by McShane et al based on 1,000 perturbations.18 The correlation between robust molecular clusters and pCR as well as classic histopathologic characteristics were examined by the 2 test. Multidimensional scaling (MDS) was performed to determine whether the expression profiles are clustered (rather than represent the same multivariate Gaussian distribution) and whether the clusters correlate with response, with the Euclidian distance as metric. Statistical analysis was performed by using the BRB-Arraytools software package (http://linus.nci.nih.gov/BRB-ArrayTools.html).
RESULTS
Sample Disposition and Patient Characteristics at INT-Milan
Ninety-five core biopsy tumor specimens were submitted for gene expression analysis by RT-PCR. Pathology review indicated that two samples had little or no tumor tissue (< 5% of the section) and were not assessable. The yield of total RNA was insufficient to assay (< 500 ng) in four patients. RT-PCR profiles were within the specifications of the assay for all patients.
Characteristics at diagnosis for the 89 assessable patients in Milan are listed in Table 1. Seventy-six patients had infiltrating ductal carcinoma. The mean tumor size at diagnosis was 6.4 cm (standard deviation, 2.3 cm).
The outcomes following primary chemotherapy are summarized in Table 2. Overall, 11 patients (12%) had a pCR.
Patients whose tumors were ER-negative by IHC analysis of the core biopsy specimen appeared to be more likely to have a pCR. Seven (23%) of the 31 patients (95% CI, 8% to 37%) with ER-negative breast cancer by IHC had a pCR. Four (8%) of the 52 patients (95% CI, 1% to 15%) with ER-positive breast cancer by IHC had a pCR. There was no correlation between patient age or tumor grade and pCR.
IHC and RT-PCR for ER and PR in INT-Milan Patients
IHC and RT-PCR measurements of tumor expression of ER and PR were compared. The concordance between RT-PCR (mRNA) and IHC (protein) was high for both ER ( = .84; 95% CI, 0.71 to 0.96) and PR ( = .71; 95% CI, 0.56 to 0.86).
Univariate Analysis of Gene Expression and pCR in INT-Milan Patients
The odds ratio for pCR for each of the candidate genes with P < .05 are given in Table 3. Thirty genes had a P value of < .01 and a total of 86 genes had a P value < .05. Simulations based on randomly assigning pCR to 11 patients indicated that four genes would be expected to have a P value of < .01 and 19 genes have a P value < .05 by chance alone.
A number of the genes that correlated with pCR fell into categories defined by biologic function, prominently: proliferation (eg, MCM6, E2F1, MYBL2), apoptosis (BBC3, BAD, DR4, TP53BP1), invasion and metastasis (FYN and MMP12), and drug resistance/metabolism (ABCC5, ALDH1A1, CYP3A4).
The quantitative expression of the ER gene by RT-PCR significantly correlated with the likelihood of pCR with an odds ratio of 0.78 (a 22% decrease in the likelihood of pCR with each increment of one unit in ER expression).
Correlated Gene Expression in INT-Milan Patients
Unsupervised cluster analysis of the 86 genes that were associated with pCR in univariate analysis was performed to identify groups of co-expressed genes that are related to the likelihood of pCR (Fig 1). Three prominent groups of such co-expressed genes were identified—an ER-related gene cluster, a proliferation-related gene cluster, and an immune-related gene cluster. Higher expression of an ER gene cluster (including PR, SCUBE2, ER, NPD009, GATA3, IGF1R, IRS1) correlated with a lower likelihood of pCR. On the other hand, higher expression of a proliferation gene cluster (including CDC20, E2F1, MYBL2, TOP2A, FBXO5, MCM2, MCM6, CDC25B) or of an immune-related gene cluster (including MCP1, CD68, CTSB, CD18, ILT-2, CD3z, FasL, HLA.DPB1, GBP1) correlated with a higher likelihood of pCR.
The fact that a substantial percentage (22%) of tested candidate genes correlated with pCR in univariate analysis is partially explained by their co-expression. Forty-six of the 86 "hits" with respect to pCR were in either the ER cluster (defined by r > 0.45, using GATA3 as reference), the proliferation cluster (defined by r > 0.45 using MCM6 as reference) or the immune response cluster (defined by r > 0.45 using ILT-2 as a reference).
On the other hand, a number of genes that did not belong to one of these three clusters of co-expressed genes also correlated with pCR. For example, higher expression of the ?-tubulin gene (TUBB), the target of paclitaxel, was significantly correlated with a lower likelihood of pCR, as was expression of the DNA repair genes MSH3 and ERCC1. Higher expression of the transcriptional regulator ID2 was correlated with higher likelihood of pCR (odds ratio, 2.37; P = .0026).
Multivariate Analysis of Gene Expression and pCR in INT-Milan Patients
Multivariate analysis was performed to identify multi-gene models for future testing. Due to the limited number of pathologically complete responders under study (11 pCRs) and the large number of genes being tested (384 genes), many multiple-gene models can discriminate pCR. For instance, a three-gene model with TBP, ILT-2 and MCM6 generated predicted probabilities of pCR that were nearly 100% concordant with observed probabilities. Consequently, bootstrap resampling,27 based on 1,000 resamplings with replacement of the original dataset, was performed to assess the arbitrariness of stepwise variable selection procedures based on the probit model. Using a 0.1 criterion for both entry and exit Wald 2 probabilities for an effect, results revealed that 26% of the selected models included only a single gene with an average concordance between predicted and observed probability of pCR greater than 90%. Furthermore, all multi-gene models contained no more than four genes and yielded an average concordance between predicted and observed probability of pCR of 97%. Consequently, multigene models should be considered exploratory and interpreted with caution.
RS and pCR in INT-Milan Patients
RS is calculated from the expression of 21 genes (16 cancer-related genes and five reference genes) as measured by the RT-PCR.17 The RS (on a scale from 0 to 100) is a function of the expression of four groups of genes plus three individual genes: an ER-related group—ER, PR, Bcl2, and SCUBE2; a proliferation group—Ki-67, MYBL2, Survivin, Cyclin B1, and STK15; a HER2 group—HER2 and GRB7; an invasion group—Stromelysin 3 and Cathepsin L2; and the three individual genes—GSTM1, BAG1, and CD68.
Probit regression,28 based on iteratively reweighted least squares, was used to assess the correlation of RS to pCR. A global likelihood ratio test comparing the (full) model with RS to the (null) model excluding RS was performed, resulting in a P value of .005. Model assessments, including assessments of goodness of fit27 (Hosmer and Lemeshow test, P = .83) and residual analysis did not reveal significant model departures.
Additionally, model validation29 was performed to confirm that the results were not spurious and to assess the performance of the RS model with respect to discrimination and potential overfitting. Bootstrap resampling,30 based on 1,000 resamplings with replacement of the original data set, yielded bias-corrected measures of model discrimination31-33 (Somer's D rank correlation, 0.50; generalized R2 = 0.14; calibrated slope, 1.13) indicating that the RS provides good discrimination of pCR. Empirical 0.5% and 99.5% percentiles for the RS regression coefficient were 0.001 and 0.045, respectively, further suggesting that the model is predictive of pCR. Predicted values of the probability of pCR as a function of RS derived from the probit model fit and normal two-sided 95% CIs are provided in Figure 2. The probability of pCR increased with RS.
Assessment of the 86 Informative Genes in a Second Independent Patient Cohort
Gene expression data from 82 patients with stage I-III breast cancer was used to assess the response discriminating value of the 86 genes identified by RT-PCR. All of these patients received sequential paclitaxel followed by FAC preoperative chemotherapy at MDACC. Characteristics at diagnosis for the 82 assessable patients treated at MDACC in Houston and their outcomes after neoadjuvant chemotherapy are given in Tables 1 and 2, respectively. Overall, 21 patients (26%) had a pCR.
Seventy-nine of the 86 genes identified as predictive in the INT-Milan cohort were represented on the Affymetrix U133A chip by 179 individual probe sets. The seven genes not on the array were AK055699, CCND1, Contig 51037, DR4, IGFBP5, KIAA1209, and RHOC.
The correlation of each probe set with pCR is shown in Table 4. Thirty-seven probe sets corresponding to 24 genes demonstrated correlation with pCR with P .05 (expected false discovery, four genes). Of the 79 genes, 32 (40%) showed association with pCR at P .1 (expected false discovery, eight genes). In all cases, the association of expression with pCR was in the same direction in the MDACC patients as had been observed in the INT-Milan patients.
We also performed supervised hierarchical clustering using all 179 probe sets to see if the cases fall into distinct molecular clusters that correspond to different rates of pCR. The resulting dendogram had two main clusters that were robust and reproducible (R-index = 0.938, D-index = 1.675; Fig 3). One cluster group with 39 cases included five pCRs (pCR rate, 12%), the other cluster had 43 cases including 16 pCRs (37%), P = .01. The cluster groups also correlated with clinical ER status assessed by immunohistochemistry (P < .001). Multidimensional scaling with all 179-probe sets also revealed that cases with pCR clustered together and the Global Test for significance yielded P < .001. These results indicate that many of the genes that were associated with pCR in the RT-PCR data from Milan had response-discriminating value in an independent data set even when measured with a different gene expression profiling platform.
DISCUSSION
We have identified a set of genes that in their expression correlate with pCR to neoadjuvant doxorubicin and paclitaxel. Of 384 candidate genes tested by RT-PCR analysis, the expression of 86 significantly correlated with pCR (P < .05). Pathologic complete response is less frequent than major clinical response, but is the independent variable most strongly associated with the likelihood of long-term disease-free survival.1,2,4
Because this study tested hundreds of candidate genes, a number of the 86 predictive genes may be false discoveries. Nevertheless, it is highly likely that many of them are valid. First, simulations that assign pCR to 11 patients at random show that at P < .05 it is expected to find only 19 positive genes by chance alone. The findings are also biologically plausible since these genes represent functional categories that are reported to influence sensitivity and resistance to chemotherapy (ie, apoptosis, invasion, metastasis, drug resistance/metabolism, proliferation, ER). Importantly, the response discriminating value of these genes was confirmed in an independent data set examined with a different gene expression profiling technology. Twenty-four (30%) of the 86 genes showed similar correlation with pCR (P < .05) in the microarray data. An additional eight genes showed modest correlation with pCR (P < .1). All of these genes had the same directional association with pCR in both cohorts of patients. Thus, many genes identified by RT-PCR had response discriminating value when measured by oligonucleotide arrays. This is notable because the two platforms have very different sensitivity and dynamic range to measure mRNA expression.34,35 Also, the tissues used for discovery in Milan were formalin-fixed, paraffin-embedded core-needle biopsies, whereas the tissues used in the confirmation analysis were obtained with FNA. These technical differences could partly explain why fewer than half of the genes showed high correlation with pCR in both data sets. The fact that a number of genes retained their predictive value in totally different groups of patients undergoing similar chemotherapy in different cancer centers seems to fulfill the type of validation recently advocated to minimize the risk of "noise" discovery.36
The 86 predictive genes included three particularly prominent co-expression clusters. Pathologic complete response correlated with higher expression of genes regulating proliferation (eg, CDC20, E2F1, MYBL2, TOPO2A) and immune response (eg, MCP1, CD68, CTSB, CD18, ILT-2, CD3z, FasL, HLA.DPB1), and with lower expression of genes related to the expression of the ER (eg, ER, PR, SCUBE2, and GATA3). These relationships of proliferation and ER to chemotherapy response are consistent with several previously reported observations.37-40
The observation that higher expression of an immune-related gene cluster is associated with chemotherapy benefit is remarkable. It is possible that the pretreatment host response may enhance the ability of chemotherapy to eliminate cancer cells. This observation should be evaluated in future studies.
It is noteworthy that the multi-drug resistance gene MDR-1, which has been frequently implicated in chemotherapy resistance in vitro,41-43 was not related to chemotherapy response here. MDR-1 also did not link to clinical anthracycline resistance in a recent paper by Faneyte et al.44 One of the genes (CYBA) previously identified by Chang et al was correlated with pCR in this study.21
Genes related to proliferation (Ki-67, Survivin, MYBL2, STK15, Cyclin B1) and the ER (ER, PGR, Bcl2, SCUBE2) are significantly represented in the 21 gene panel employed by the gene expression–based RS assay.17 The RS has been validated to quantify the risk of recurrence in tamoxifen-treated patients with node-negative, ER-positive breast cancer. We show here that RS strongly correlated with pCR. This has a provocative clinical implication, namely, that patients with high RS values, who are most likely to experience recurrence,17 are the very patients most likely to receive the greatest clinical benefit from chemotherapy treatment. It will be important to further evaluate this concept in future studies.
In conclusion, this study correlated the expression of certain genes to the probability of pathologic complete response. The observation that many of the discriminating genes maintained their predictive value across two cohorts and two different gene expression profiling platforms makes the findings compelling enough to subject these genes to further clinical evaluation. The results of this study are consistent with a growing body of data that indicates that not all breast cancer patients respond equally to chemotherapy. Successful prediction of a durable benefit of chemotherapy using routinely obtained fixed tumor tissue would make possible and widely applicable the tailoring of toxic chemotherapy to individual patients.
Authors' Disclosures of Potential Conflicts of Interest
Although all authors completed the disclosure declaration, the following authors or their immediate family members indicated a financial interest. No conflict exists for drugs or devices used in a study if they are not being evaluated as part of the investigation. For a detailed description of the disclosure categories, or for more information about ASCO's conflict of interest policy, please refer to the Author Disclosure Declaration and the Disclosures of Potential Conflicts of Interest section in Information for Contributors.
Acknowledgment
We acknowledge John Morlan for his assistance with RNA extraction, Chithra Sangli for biostatistical input, and Michael Kiefer for help with candidate gene selection.
NOTES
Supported in part by a grant from Genomic Health Inc, Redwood City, CA.
Presented in abstract form at the 40th Annual Meeting of American Society of Clinical Oncology, New Orleans, LA, June 5-8, 2004.
Terms in blue are defined in the glossary, found at the end of this issue and online at www.jco.org.
Authors' disclosures of potential conflicts of interest are found at the end of this article.
REFERENCES
Fisher B, Bryant J, Wolmark N, et al: Effect of preoperative chemotherapy on the outcome of women with operable breast cancer. J Clin Oncol 16:2672-2685, 1998
Bonadonna G, Valagussa P, Brambilla C, et al: Primary chemotherapy in operable breast cancer: Eight-year experience at the Milan Cancer Institute. J Clin Oncol 16:93-100, 1998
Smith IC, Heys SD, Hutcheon AW, et al: Neoadjuvant chemotherapy in breast cancer: Significantly enhanced response with docetaxel. J Clin Oncol 20:1456-1466, 2002
Kaufmann M, von Minckwitz G, Smith R, et al: International expert panel on the use of primary (preoperative) systemic treatment of operable breast cancer: Review and recommendations. J Clin Oncol 21:2600-2608, 2003
Rouzier R, Wagner P, Morandi P, et al: Gene expression profiling of primary breast cancer. Curr Oncol Rep 7:38-44, 2005
Sorlie T, Perou CM, Tibshirani R, et al: Gene expression patterns of breast carcinomas distinguish tumor subclasses with clinical implications. Proc Natl Acad Sci USA 98:10869-10874, 2001
van de Vijver MJ, He YD, van't Veer LJ, et al: A gene-expression signature as a predictor of survival in breast cancer. N Engl J Med 347:1999-2009, 2002
Chang J, Powles TJ, Allred DC, et al: Biologic markers as predictors of clinical outcome from systemic therapy for primary operable breast cancer. J Clin Oncol 17:3058-3063, 1999
Sotiriou C, Powles TJ, Dowsett M, et al: Gene expression profiles derived from fine needle aspiration correlate with response to systemic chemotherapy in breast cancer. Breast Cancer Res 4:R3, 2002
Assersohn L, Gangi L, Zhao Y, et al: The feasibility of using fine needle aspiration from primary breast cancers for cDNA microarray analyses. Clin Cancer Res 8:794-801, 2002
Ayers M, Symmans WF, Stec J, et al: Gene expression profiles predict complete pathologic response to neoadjuvant paclitaxel and fluorouracil, doxorubicin, and cyclophosphamide chemotherapy in breast cancer. J Clin Oncol 22:2284-2293, 2004
Ellis M, Davis N, Coop A, et al: Development and validation of a method for using breast core needle biopsies for gene expression microarray analyses. Clin Cancer Res 8:1155-1166, 2002
Karsten SL, Van Deerlin VM, Sabatti C, et al: An evaluation of tyramide signal amplification and archived fixed and frozen tissue in microarray gene expression analysis. Nucleic Acids Res 30:E4, 2002
Ramaswamy S, Golub TR: DNA microarrays in clinical oncology. J Clin Oncol 20:1932-1941, 2002
Cronin M, Pho M, Dutta D, et al: Measurement of gene expression in archival paraffin-embedded tissues: Development and performance of a 92-gene reverse transcriptase-polymerase chain reaction assay. Am J Pathol 164:35-42, 2004
Specht K, Richter T, Muller U, et al: Quantitative gene expression analysis in microdissected archival formalin-fixed and paraffin-embedded tumor tissue. Am J Pathol 158:419-429, 2001
Paik S, Shak S, Tang G, et al: A multigene assay to predict recurrence of tamoxifen-treated, node-negative breast cancer. N Engl J Med 351:2817-2826, 2004
McShane LM, Radmacher MD, Simon R: Methods for assessing reproducibility of clustering patterns observed in analyses of microarray data. Bioinformatics 18:1462-1469, 2002
Perou CM, Sorlie T, Eisen MB, et al: Molecular portraits of human breast tumours. Nature 406:747-752, 2000
van't Veer LJ, Dai H, van de Vijver MJ, et al: Gene expression profiling predicts clinical outcome of breast cancer. Nature 2002;415(6871):530-6.
Chang JC, Wooten EC, Tsimelzon A, et al: Gene expression profiling for the prediction of therapeutic response to docetaxel in patients with breast cancer. Lancet 362:362-369, 2003
Duran GE, Wang YC, Francisco B, et al: Identification of genes associated with non-MDR1 taxane resistance in human ovarian carcinoma cell lines by microarray analysis. Am Assoc Cancer Res, 2003 (abstr 5760)
Lamendola DE, Duan Z, Yusuf RZ, et al: Molecular description of evolving paclitaxel resistance in the SKOV-3 human ovarian carcinoma cell line. Cancer Res 63:2200-2205, 2003
Duan Z, Feller AJ, Penson RT, et al: Discovery of differentially expressed genes associated with paclitaxel resistance using cDNA array technology: analysis of interleukin (IL) 6, IL-8, and monocyte chemotactic protein 1 in the paclitaxel-resistant phenotype. Clin Cancer Res 5:3445-3453, 1999
Elston CW, Ellis IO: Pathological prognostic factors in breast cancer, I: The value of histological grade in breast cancer: experience from a large study with long-term follow-up. Histopathology 19:403-410, 1991
Symmans WF, Ayers M, Clark EA, et al: Total RNA yield and microarray gene expression profiles from fine-needle aspiration biopsy and core-needle biopsy samples of breast carcinoma. Cancer 97:2960-2971, 2003
Hosmer DW, Lemeshow S: Applied Logistic Regression. New York, NY, Wiley, 1989
B. J. T. Morgan: Analysis of Quantal Response Data. New York, NY, Chapman and Hall, 1992
Harrel FE Jr: Regression Modeling Strategies: With Applications to Linear Models, Logistic Regression, and Survival Analysis. New York, NY, Springer-Verlag, 2001
Efron B, Tibshirani R: An Introduction to the Bootstrap. New York, NY, Chapman and Hall, 1993
Somers RH: A new asymmetric measure of association for ordinal variables. Am Sociol Rev 27:799-811, 1962
Nagelkerke NJD: A note on a general definition of the coefficient of determination. Biometrika 78:691-692, 1991
Efron B: Estimating the error rate of a prediction rule: Improvement on cross-validation. J Am Stat Assoc 78:316-331, 1983
Kuo WP, Jenssen TK, Butte AJ, et al: Analysis of matched mRNA measurements from two different microarray technologies. Bioinformatics 18:405-412, 2002
Yuen T, Wurmbach E, Pfeffer RL, et al: Accuracy and calibration of commercial oligonucleotide and custom cDNA microarrays. Nucleic Acids Res 30:e48, 2002
Ioannidis JP: Microarrays and molecular research: Noise discovery? Lancet 365:454-455, 2005
Lippman ME, Allegra JC, Thompson EB, et al: The relation between estrogen receptors and response rate to cytotoxic chemotherapy in metastatic breast cancer. N Engl J Med 298:1223-1228, 1978
Kuerer HM, Newman LA, Smith TL, et al: Clinical course of breast cancer patients with complete pathologic primary tumor and axillary lymph node response to doxorubicin-based neoadjuvant chemotherapy. J Clin Oncol 17:460-469, 1999
Colleoni M, Minchella I, Mazzarol G, et al: Response to primary chemotherapy in breast cancer patients with tumors not expressing estrogen and progesterone receptors. Ann Oncol 11:1057-1059, 2000
Gianni L, Baselga J, Eiermann W, et al: First report of the European Cooperative Trial in operable breast cancer (ECTO): Effects of primary systemic therapy (PST) on local-regional disease. Proc Am Soc Clin Oncol, 2002 (abstr 132)
Penson RT, Oliva E, Skates SJ, et al: Expression of multidrug resistance-1 protein inversely correlates with paclitaxel response and survival in ovarian cancer patients: A study in serial samples. Gynecol Oncol 93:98-106, 2004
Wu H, Hait WN, Yang JM: Small interfering RNA-induced suppression of MDR1 (P-glycoprotein) restores sensitivity to multidrug-resistant cancer cells. Cancer Res 63:1515-1519, 2003
Burger H, Foekens JA, Look MP, et al: RNA expression of breast cancer resistance protein, lung resistance-related protein, multidrug resistance-associated proteins 1 and 2, and multidrug resistance gene 1 in breast cancer: Correlation with chemotherapeutic response. Clin Cancer Res 9:827-836, 2003
Faneyte IF, Kristel PM, van de Vijver MJ: Multidrug resistance associated genes MRP1, MRP2 and MRP3 in primary and anthracycline exposed breast cancer. Anticancer Res 24:2931-2939, 2004(Luca Gianni, Milvia Zambe)