Breast Cancer. Molecular Classification of Breast Cancer: Limitations and Potential. The Oncologist 2006;11:

Transcription

1 Breast Cancer Molecular Classification of Breast Cancer: Limitations and Potential Lajos Pusztai, a Chafika Mazouni, a Keith Anderson, b Yun Wu, c W. Fraser Symmans c Departments of a Breast Medical Oncology, b Bioinformatics and Applied Mathematics, and c Pathology, University of Texas M.D. Anderson Cancer Center, Houston, Texas, USA Key Words. Breast cancer Molecular profiling Molecular classification DNA microarrays Learning Objectives After completing this course, the reader will be able to: 1. Discuss the technical variables that affect accuracy of estrogen receptor immunohistochemistry. 2. Describe emerging methods to quantify estrogen receptor expression and predict the prognosis of estrogen receptor-positive patients. 3. Discuss limitations of current gene expression-based molecular classification of breast cancer. 4. Explain the conceptual differences between unsupervised molecular class discovery methods and supervised clinical outcome prediction models such as multigene prognostic signatures. 5. Interpret results of DNA microarray literature as they relate to diagnosis and prognosis of breast cancer. CME Access and take the CME test online and receive 1 AMA PRA Category 1 Credit at CME.TheOncologist.com Abstract Reverse transcription polymerase chain reaction and DNA microarrays are increasingly used in the clinic and in clinical research as prognostic or predictive tests. Results from these tests led to novel risk stratification methods and to new molecular classification of breast cancer. Some of these tools already complement existing diagnostic tests and can aid medical decision making in some situations. Better understanding of the molecular classes of breast cancer, independent of their prognostic and predictive values, may also lead to new biological insights and eventually to better therapies that are directed toward particular molecular subsets. However, there is substantially less experience with these emerging technologies than with the more established methods, the accuracy of which is often overestimated. This review discusses some of the limitations and strengths of current gene expression-based molecular classification of breast cancer. To provide context for this discussion, we also briefly examine the performance of estrogen receptor immunohistochemistry, which represents an essential part of the routine diagnostic workup for all breast cancer patients. The Oncologist 2006;11: Molecular Classification of Breast Cancer Breast cancer is a clinically heterogeneous disease, and existing histological classifications do not fully capture the varied clinical course of this disease. Histological type, grade, tumor size, lymph node involvement, and estrogen receptor α (ER) and HER-2 receptor status all influence prognosis and the probability of response to systemic therapies. These clinical variables can be combined into Correspondence: Lajos Pusztai, M.D., D.Phil., University of Texas M.D. Anderson Cancer Center, Department of Breast Medical Oncology, Unit 1354, PO Box , Houston, Texas , USA. Telephone: ; Fax: ; lpusztai@mdanderson. org Received January 19, 2006; accepted for publication July 10, AlphaMed Press /2006/$20.00/0 doi: / theoncologist The Oncologist 2006;11:

2 Pusztai, Mazouni, Anderson et al. 869 multivariate outcome prediction models. The Nottingham Prognostic Index and Adjuvant! Online ( adjuvantonline.com) represent two of the probably most commonly used prognostic tools [1, 2]. A similar tool was recently proposed to predict probability of response to preoperative chemotherapy ( care_centers/breastcenter/dindex.cfm?pn=448442b2-3ea5-4bac a9553e63) [3]. However, regardless of what outcome prediction model is used, there remains substantial variability in disease outcome within each risk category. The ability to predict a 10% or 30% risk for recurrence is useful in medical decision making; however, it is also an admission that we cannot reliably predict the true binary outcome, recurrence versus no recurrence, at the level of the individual. It is generally accepted that the different clinical courses of patients with histologically identical tumors is a result of molecular differences among cancers. It follows that detailed molecular analysis of the cancer could yield information that will improve prognostic prediction. Every molecular analytical method that can be applied to human cancer tissue has potential to become a prognostic/predictive test. Naturally, there is greater experience with older methods, and we understand their diagnostic value better. Immunohistochemistry (IHC) was developed more than 25 years ago and currently forms the cornerstone of molecular classification of breast cancer into ER-positive and ER-negative categories. Nucleic acid in situ hybridization was introduced more than 15 years ago and is now also routinely used to classify breast cancer into HER-2 amplified or nonamplified categories. A large number of other single gene molecular markers were also assessed in the past 25 years using these technologies but failed to establish themselves in the clinic for various reasons [4]. More recently, high throughput proteomic and gene-expression profiling methods are being explored as diagnostic tools. The basic premise of these emerging tests is that they simultaneously quantify the expression of multiple genes and combine the gene expression measurements into prediction scores that may foretell clinical outcome more accurately than any of the genes alone. Classification Based on ER Status Determination of ER status is an essential part of the diagnostic workup of all breast cancer patients. The information is most useful to determine whether a patient is a candidate for endocrine therapy or not. ER-negative patients do not benefit from endocrine interventions. The current gold standard to asses ER status is IHC performed on formalinfixed, paraffin-embedded cancer tissue. This diagnostic test is routinely used in the clinic, and major therapeutic decisions depend on the results; however, its reliability is far from perfect. The existing IHC assays have only modest positive predictive value (30% 60%) for response to single-agent hormonal therapies [5, 6]. Furthermore, there is substantial intra- and interlaboratory variation in ER results because fixation, antigen retrieval, and staining methods may differ among laboratories [7 9]. Interpretation of staining can also be subjective to the individual pathologist or the threshold setting of the image analysis system being used [10]. In one study, 200 clinical laboratories received sections from the same three tumors that showed low, moderate, or high ER expression in a reference laboratory. Each laboratory performed its own IHC assessment, and the results indicated considerable interlaboratory variation [7]. The false-negative rates were as high as 30% 60% (depending on the cutoff) in the low ER-positive case (Table 1A). In another report, the effect of the duration of formalin fixation on ER staining was examined. Twentyfour large, strongly ER-positive tumors were sliced, and pieces of the tumors were fixed in formalin for 3, 6, 8, or more hours separately [9]. The mean ER IHC scores, on a 0 7 scale, were 2.5 for blocks fixed for 3 hours, 5.75 for blocks fixed for 6 hours, and 6.7 for blocks fixed for 8 hours. Some strongly ER-positive tumors were completely negative when the shortest fixation time was used (Table 1B). The method and duration of antigen retrieval also affect IHC results. It is important to be aware of the substantial technical variability in quantifying ER by IHC, and when a false-negative result is suspected based on the clinical characteristics of a case, a repeat biopsy may be in order. The technical variability and discrepancies in reproducibility of results among laboratories is not limited to ER IHC. Substantial discordance among HER-2 results generated in different laboratories from the same specimen has also been reported. For example, the level of concordance for HER-2 IHC results was 80%, and for HER-2 fluorescence in situ hybridization it was 85% when the same specimens were tested in local and central laboratories [11, 12]. The modest positive predictive value of current methods and the variable reproducibility of results fuel efforts to develop more accurate and more reliable predictors of benefit from hormonal therapy. There are some data to suggest that the higher the level of expression of ER, the higher the probability of benefit from endocrine therapy [13]. Therefore, more accurate quantification of ER expression over a broad dynamic range may alone represent an improvement over the semiquantitative ER IHC. There are now several methods that can reliably measure ER mrna expression over a broad range, foremost among these are quantitative reverse transcription polymerase chain reaction (RT-PCR) and DNA microarrays. Since ER protein and ER mrna

3 870 Molecular Classification of Breast Cancer Table 1. Technical variability of estrogen receptor (ER) immunohistochemistry (IHC) A. Interlaboratory reproducibility of ER IHC results across 200 laboratories Three reference cases IHC Result (1) High ER (2) Intermediate ER (3) Low ER 10% ER-positive cells 99% 84.5% 37% 1% ER-positive cells 99.5% 88% 66% B. Impact of formalin fixation time on ER score assessed by IHC Fixation time Mean IHC score (range) 3 hours 2.46 (0 6) 6 hours 5.75 (2 7) 8 hours 6.7 (5 7) Panel A shows the proportion of 200 different clinical laboratories that reported 10% or 1% ER positivity for the same three reference tumors that represented high, intermediate, and low ER expression, respectively. Adapted from [7]. Panel B shows results from an experiment in which pieces of 24 ER-positive tumors were fixed for a variable length of time before ER IHC was performed. IHC scores are shown as a function of fixation time. Adapted from [9]. levels correlate closely, it is reasonable to hypothesize that determination of ER status by mrna measurements can be of clinical value [14]. Oncotype DX and Other Prediction Scores The mere presence of ER does not guarantee functional activity, and other molecular events, unrelated to ER signaling, may also influence sensitivity to hormonal therapy. Oncotype DX (Genomic Health Inc., Redwood City, CA) represents an important conceptual advance in the diagnosis of ER-positive breast cancers. This RT-PCR-based assay not only measures ER mrna expression in a highly quantitative and reproducible manner, but it also measures the expression of several downstream ER-regulated genes (PR, BCL-2, SCUBE-2) that may contain information on ER functionality. The same assay also quantifies HER-2 expression and proliferation-related genes [15]. Combining information from each of these measurements into a single prediction score can provide a superior method of outcome prediction than ER IHC alone. A study examined the correlation between the Oncotype DX recurrence score and the likelihood of distant relapse in 668 ER-positive, nodenegative, tamoxifen-treated patients who were enrolled in the National Surgical Adjuvant Breast and Bowel Project (NSABP) clinical trial B14 [15]. Fifty-one percent, 22%, and 27% of these ER-positive patients were categorized as low, intermediate, and high risk for recurrence after tamoxifen therapy, respectively. The observed 10-year distant recurrence rates were 6.8%, 14.3%, and 30.5% in the three risk categories, respectively (p < 0.001). In a multivariate analysis, the recurrence score predicted relapse and overall survival independently of age and tumor size. Similar results were observed in a community-based patient population in a separate study [16]. A recent report examined the value of the recurrence score for predicting benefit from adjuvant cyclophosphamide, methotrexate, and 5-fluorouracil (CMF) chemotherapy in 651 patients with ER-positive, node-negative breast cancer included in the NSABP B20 randomized study [17]. Higher recurrence scores were associated with greater benefit from adjuvant CMF chemotherapy (test for interaction, p =.038). The hazard ratio for distant recurrence after CMF chemotherapy was 1.31 (95% confidence interval [CI], ) for patients with a recurrence score <18 and 0.26 (95% CI, ) for patients with a recurrence score >31. The absolute improvement in 10-year distant recurrence-free survival was 28% (60% vs. 88%) in patients with a recurrence score >31, while there was no benefit in patients with a recurrence score <18. In that study, neither the quantity of ER measured by ligand-binding assay nor the tumor grade was predictive of benefit from chemotherapy. These data indicate that a high recurrence score could identify a subset of women with ER-positive and node-negative breast cancer that have a high risk for recurrence with tamoxifen therapy alone, independent of grade and quantitative ER status, and this risk can be reduced with the administration of adjuvant CMF chemotherapy. Several other multigene prediction scores have been developed in order to refine the predictive value of ER and better estimate who will or will not do well with endocrine therapy. In one study, 81 genes were found to be differentially expressed between tamoxifen-sensitive and tamoxifen-resistant, ER-positive breast cancers (n = 46), and a 44- gene predictive signature was developed from the data. This signature was tested on 66 independent ER-positive cases and could select patients who had prolonged progressionfree survival (odds ratio, 3.16) [18]. In another report, a 200- gene ER reporter index was developed from genes whose The Oncologist

4 Pusztai, Mazouni, Anderson et al. 871 expression was highly associated with the ER. This reporter index, which did not include the ER itself, could accurately determine clinical ER status in independent cases and predict recurrence-free survival in a small set of patients who received adjuvant tamoxifen [19]. These genomic scores are in the early stages of development and have not yet undergone the same degree of clinical validation as the Oncotype DX assay. However, these data together suggest that mrna-based ER measurements provide a better quantification of ER expression over a broader dynamic range than ER IHC. Figure 1A shows ER mrna expression measurements by the Affymetrix U133A GeneChip array (Affymetrix, Inc., Santa Clara, CA) (n = 133) in four ER categories of patients defined by IHC, including: IHC negative (0% positive cells by IHC), weakly positive (<10% positive cells), positive (11% 50% positive cells), and strongly positive (>50% positive cells). This plot was generated from data that are available through a previous publication [20]. Figure 1B illustrates the reproducibility of ER mrna measurements in 31 between- and within-laboratory replicate experiments. Data are based on previously published results [21]. Inclusion of other genes that reflect ER activity, HER- 2 expression, and proliferative status of the tumor can substantially improve the predictive value of ER (i.e., further stratify ER-positive patients by outcome). Whereas these methods provide more readily quantifiable results, they too remain susceptible to variations in initial sample handling. Both proteins and RNA can rapidly change and degrade if the surgical specimen or diagnostic needle biopsy are not handled properly. Figure 1. Estrogen receptor α (ER) mrna measurements by Affymetrix U133A GeneChip array in 133 fine-needle biopsies of breast cancer. (A): ER mrna expression (ESR1 probe _at) is shown in four cohorts of patients corresponding to cases with 0%, 1% 10%, 11% 50%, and >50% ER-positive cells by routine immunohistochemistry. Data from [20]. (B): Reproducibility of ER mrna measurements in 31 replicate experiments. The same RNA samples were profiled in replicates in two different laboratories or in the same laboratory several months apart Data from [21]. Gene Expression-Based Molecular Classes: Luminal-like, Basal-like, Normal-Like, and HER-2-Positive Classification of breast cancer into ER-positive and HER-2- positive categories is mainly useful because targeted therapies are available that are only effective in these subsets of patients. The advent of high throughput gene-expression profiling technologies allowed investigators to ask a broader question. How many different molecular subtypes of breast cancer exist based on complex mrna expression patterns? It might be useful to identify the various molecular classes because the different subsets may have different natural histories. Also, the molecular signatures that define particular groups may lead to the discovery of new therapeutic targets and treatments that are effective in particular molecular subsets. The first study to examine comprehensive gene-expression patterns of breast cancer suggested that at least four major molecular classes of breast cancer exist: luminal-like, basal-like, normal-like, and HER-2 positive [22]. Subsequent studies confirmed that there are large-scale gene-expression differences between ER-positive (mostly luminal-like) and ER-negative (mostly basallike) cancers and suggested that further molecular subsets also exist [23 25]. The prognosis and chemotherapy sensitivity of the different molecular subgroups are different. Luminal-like cancers tend to have the most favorable long-term survival compared with the others, whereas basal-like and HER-2-positive tumors are more sensitive to chemotherapy [26, 27]. When interpreting these observations, it is important to keep in mind that many of these correlations are expected because of the strong association between molecular class and conventional histopathologic variables. For example, in one experiment, all luminal-like cancers were ER positive and 63% of them were also low or intermediate grade, whereas 95% of basal-like cancers were ER negative, and 91% of them were high grade [27]. These associations may explain the seemingly contradictory observation that basal-like cancers have a worse prognosis than luminal-like cancers in spite of their higher sensitivity to chemotherapy. Luminal-like cancers are ER positive and lower grade, and therefore sensitive to endocrine therapy, and may have a more favorable prognosis than the ER-negative and high-grade basal-like cancers, even in the absence of any therapy. Limitations of Current Molecular Classification Despite these promising results, there are major limitations in our ability to consistently assign a molecular class to new cases of breast cancer. Foremost among these is the lack of a standardized molecular class prediction method. It also remains unknown just how many molecular classes

5 872 Molecular Classification of Breast Cancer of breast cancer there are. In fact, a more practical question is not How many true molecular classes exist? but How many classes can be reliably identified with the currently available data? Hierarchical clustering, multidimensional scaling, self-organizing maps are some of the most commonly used methods to discover and display structure in high-dimensional gene-expression data. Hierarchical clustering algorithms group samples together based on similarity in their patterns of gene expression. The results are displayed as a dendrogram; the more similar the cases, the closer they are located on the end branches of the dendrogram tree. The longer the arm of the branch, the greater the difference in gene-expression pattern. This class discovery method was applied to data from three normal breast samples and 40 different breast tumors, including 20 repeated measurements from the same tumor to discover molecular classes of breast cancer [22]. The investigators selected the genes that showed the greatest between-sample variability, and using these 1,753 genes in hierarchical clustering, they observed two main branches (clusters) and additional smaller secondary clusters within the larger branches of the dendogram. These data led to a four-way classification of breast cancer: luminal-like (expressing luminal cytokeratins 8/18), basal-like (characterized by cytokeratins 5/17), HER-2 positive (mostly, but not all, HER-2 amplified), and normal-like. Subsequently, the gene list was modified, and 476 genes, termed the intrinsic gene set, were used in a follow-up study including 78 cancers that suggested three additional subgroups within the luminal-like category, luminal-a, luminal-b, and luminal-c [24]. In a third study from the same investigators, a further modified version of the intrinsic gene list, now including 534 genes, was applied to 115 cases. The basal-like, normal-like, HER-2- positive, and two categories of luminal-like cancers were observed again [26]. The most recent publication proposes yet another gene list that is now extended to include 1,300 genes that may be applied for hierarchical clustering across multiple platforms [28]. It is reassuring that broadly similar molecular classes emerged from these series of studies; however, the rather subjective nature of class assignment based on hierarchical clustering must be acknowledged. Clustering algorithms always detect clusters, even in random data, and it is not easy to define how many true clusters a particular dendrogram tree has. For example, does Figure 2 show two or three or four or more true clusters (i.e., molecular classes) of breast cancer? Relating a seemingly apparent cluster to known clinical/biological characteristics and a distinct clinical course (e.g., luminal-like cluster roughly equivalent to ER-positive disease) is one way to validate the observation. However, the hope is that the cluster analysis will discover new, previously unrecognized classes of the disease. Unfortunately, it is easy to misinterpret dendrogram results, particularly when novel molecular subsets are proposed without some objective measure of robustness and reproducibility. McShane et al. [29] proposed a two-step approach in which a global test for clustering is performed and, if a pattern is suggested, the strength of cluster-specific reproducibility is assessed by adding Gaussian noise to the data. Application of this test to the data shown in Figure 2 yielded reproducibility indices (R) of 0.82 (luminal-like), 0.76 (normal-like), 0.85 (basal-like), and 0.78 (HER-2 positive) for the four molecular classes, which indicates only moderately stable clusters. This figure is reprinted from a previous publication [27]. Attempts to define further subsets within the four major clusters yielded very low reproducibility. A number of other statistical methods are also available to detect the optimal number of robust clusters in microarray data [30, 31]. It is important that some of these objective measures be used to interpret patterns of clustering, even if they tend to give rather conservative results. Clustering results also depend on the gene set that is used. Figure 3 illustrates how the dendrogram shown in Figure 2 changed after a different gene set, including the 1,300 highly variably expressed genes from [28], was used for clustering of the same data (n = 82 cases). The previously uniform, but least robust (R = 0.76), normal-like group is now completely dispersed across multiple branches. Five of the 22 basal-like cancers (23%) also reclustered with different molecular subsets, including three that now fall within the luminal-like category. This illustrates that it is important to use a standard gene set, otherwise clustering yields different results even for the same data. Clustering algorithms work by linking the two most similar specimens together first, and then successively merging other specimens in order of similarity. Therefore, when new cases are added to an existing data set, the previous order of clustering is revised, and a completely new dendrogram is generated. Figure 4 illustrates what happened to the dendrogram presented in Figure 2 after 51 new cases were added to the data (same gene set is used for clustering). The HER-2-positive (64% HER-2 amplified), luminal-like (91% ER positive), and basal-like (77% ER negative) groups remain distinguishable, but again the previously normallike group is no longer apparent. Based on these results, it is entirely plausible that the normal-like category does not exist as a separate group. However, it is also possible that if a set of normal breast samples were included in the data, these tumors might still cluster with the normal samples. However, because the gene-expression profile of normal breast tissue is profoundly different from that of breast cancer, by inference, tumors that are similar to normal must The Oncologist

6 Pusztai, Mazouni, Anderson et al. 873 Figure 2. Complete linkage hierarchical clustering of 82 cases with the 689 intrinsic genes. Hierarchical clustering was performed on normalized gene-expression data using 1 Pearson correlation coefficient. The four major clusters are color coded and represent the molecular classes HER-2 positive (brown), luminal-like (blue), normal-like (green), and basal-like (yellow). From Rouzier R, Perou CM, Symmans WF et al. Breast cancer molecular subtypes respond differently to preoperative chemotherapy. Clin Cancer Res 2005;11: , with permission. Figure 3. Complete linkage hierarchical clustering of 82 cases with the 1,300-gene revised intrinsic genes. Hierarchical clustering results of the same 82 cases as in Figure 1 using a revised intrinsic gene set corresponding to 1,300 genes [28]. Cases are color coded by the original molecular class as shown in Figure 1.

7 874 Molecular Classification of Breast Cancer Figure 4. Complete linkage hierarchical clustering of 133 cases with the 689 intrinsic genes. Hierarchical clustering of 133 cases, including the original 82 cases from Figure 1 plus 51 new cases, using the same intrinsic gene set. The original 82 cases are color coded by molecular class as shown in Figure 1. Raw microarray data are from [20]. also look rather different from the rest of the cancers. Fine subdivision of a dendrogram tree with a relatively small sample size is quite susceptible to producing results that are irreproducible. Even though it is tempting to subdivide the three large branches of Figure 4 that correspond to the HER-2-positive, luminal-like, and basal-like tumors into smaller subgroups, particularly because there is substantial clinical heterogeneity within these categories, it requires a substantially larger sample size to reliably define smaller molecular subsets of breast cancer. It is beyond the scope of the current review to discuss the numerous other variables, including data input, choice of distance metric, and linkage, that all have a profound effect on the shape of a dendrogram that is generated by hierarchical clustering. Rather, the purpose is to draw attention to the need to develop a standardized method and substantially large databases to define molecular classes of breast cancer. Larger data sets are required in order to develop a predictor that could prospectively assign a molecular class to new cases. Important efforts were recently made to develop such single sample class predictors [28]. Only after a standard method for class prediction that includes the platform, data normalization, gene set, and prediction rules is developed can the value of molecular classification be tested appropriately in clinical trials. Supervised Class Predictors and Prognostic and Predictive Gene Signatures A review of emerging molecular classification methods of breast cancer would not be complete without a brief discussion of gene signatures predictive of prognosis or response to chemotherapy. The fundamental difference between these and the previously discussed molecular classification method is that multigene signatures are derived through supervised class prediction methods that take into account the known clinical outcome of cases during the development of the predictor. Instead of asking the question, What molecular classes of breast cancer exist, and do these classes correspond to different clinical courses? one could ask, What gene expression differences exist between cancers that recurred and those that did not (or between tumors that responded to chemotherapy and those that did not)? Once these differentially expressed genes are identified, they can be combined into a score to predict outcome of new cases. This approach to developing a clinical outcome predictor may be applied to all cases or to subsets of cancers The Oncologist

8 Pusztai, Mazouni, Anderson et al defined by ER status or molecular class. Whether it is better to develop predictors from all samples or from particular subsets of breast cancer is an active area of research. There are examples of successful application of both approaches. A 70-gene prognostic predictor was developed from 78 breast cancers that were selected to include an even number of cases who relapsed in 5 years and who did not, but ER status and other clinical variables were not considered when the molecular predictor was developed [32]. This general prognostic predictor worked well in validation in independent cases [33, 34]. Other investigators took a different approach and identified genes that were associated with relapse separately for the ER-negative and ER-positive subsets of patients. The markers that were selected from each group were then combined to form a single 76-gene prognostic signature [35]. This predictor also performed well when tested on 180 independent cases [36]. It is intuitively and clinically appealing to develop supervised clinical outcome predictors for subsets of patients, including those with ER-negative disease or with a particular molecular class. However, this strategy also leads to additional technical and statistical challenges. The more homogeneous the patient population, the more difficult it may be to define transcriptional differences that separate the two outcome groups. For example, we developed a multigene predictor of pathologic complete response (pcr) to preoperative paclitaxel plus 5-fluorouracil, doxorubicin, and cyclophosphamide chemotherapy. When all breast cancer patients, regardless of ER status or grade, were included in the predictor development process, many differentially expressed genes could be identified and strong pcr predictors could be developed [20, 37]. However, when the same methods were applied to identify informative genes for pcr within the ER-negative subset, the false discovery rates were high as a result of the relative molecular homogeneity of ER-negative cases. Figure 5 illustrates the distribution of p-values for >20,000 genes obtained with two-sample t-statistics in three distinct data sets: randomly perturbed data (Fig. 5A), a cohort of unselected ER-positive and ERnegative breast cancers (n = 82) (Fig. 5B), and the ER-negative subset from cohort B (n = 36) (Fig. 5C). Beta uniform mixture analysis of the p-values, a statistical method used to calculate false discovery rates corresponding to particular p-values when multiple comparisons are performed [38], reveals that the p-values are not distributed evenly in the unselected breast cancer population. Rather, there is a subset of genes with small p-values and corresponding low false discovery rates (Fig. 5B). This subset of small p-values is not seen in the random data or in the data from the ERnegative cases (Fig. 5A, C). However, this observation does not necessarily indicate that there are no real transcriptional differences between cases with pcr and residual disease in the ER-negative subset. It may be that there are only a few genes with moderately large differential expression and, after correcting for multiple hypothesis testing, no gene may meet the threshold for statistical significance because of the noise from numerous variably expressed genes within the group. Analysis at the single gene level may also miss small expression differences in a larger number of genes that belong to important biological pathways. In some situations, a 20% coordinated change in the expression values of many genes in a particular metabolic pathway can have robust functional consequences [39]. These differences would not be identified easily by pairwise gene expression comparisons, such as two-sample t-statistics. Different mathematical tools, including gene set enrichment analysis, are available to test for these potentially relevant but small transcriptional differences [40]. How to integrate these more subtle transcriptional differences into diagnostic marker development remains to be learned. Future Directions Breast cancer is not one disease but a collection of several biologically different diseases. There are large-scale molecular differences between ER-positive and ER-negative cancers that reach far beyond the presence or absence of ER. HER-2-amplified cancers also have a very distinct geneexpression signature that also extends to some cancers with no apparent HER-2 gene amplification. Further molecular categories are likely to exist in addition to and within the Figure 5. Distribution of the p-values of differentially expressed genes in three separate data sets. Beta uniform mixture analysis of p-values from t-statistics comparing cases with pathologic complete response and residual disease in: randomly perturbed data (A), unselected breast cancer (B), and estrogen receptor-negative breast cancer (C). Raw microarray data are from [20].

9 876 Molecular Classification of Breast Cancer three most apparent molecular classes. Transcriptional profiling technologies can measure the expression of a large number of genes simultaneously and the combined expression values of many genes can define molecular classes more accurately than conventional single gene markers do. Some multigene prognostic predictors are already commercially available and can provide useful information for a select group of patients. However, with new technologies, new challenges also arise. Intra- and interlaboratory reproducibility of results must be determined, and the standardization of methodology has to be established. The sensitivity of prediction results to tissue acquisition methods, sample handling, and experimental noise also needs to be defined as part of the clinical evaluation process. These are relatively straightforward technical hurdles that could be solved readily by commercial and academic laboratories. A more difficult challenge may be to create a clinical infrastructure that is streamlined for molecular marker discovery and validation. This may require shifting emphasis from merely demonstrating an association between a marker and clinical outcome to proving clinical benefit from the use of a new test. In academic institutions, biomarker discovery and validation are often performed as ancillary, correlative studies on tissues that become available from patients included in therapeutic clinical trials. Therapeutic trials are invariably designed with the optimal sample size in mind to demonstrate an expected outcome, whereas the sample size of biomarker studies is often driven by availability of tissues. It might be useful to think about the development of various emerging genomic tests as a process that is conceptually similar to therapeutic drug evaluation. In this context, the goal of a phase I pharmacogenomic discovery study is to define the predictive gene set, establish the prediction rules, and determine assay cutoffs in a well-defined patient population. During phase II testing, the performance of the predictor is assessed on independent cases, and the reproducibility and robustness of the assay are determined. These initial phases of marker development may be performed on archived specimens with a prospectively defined analysis plan. The final phase of marker evaluation would include a prospectively conducted randomized clinical trial to show that the use of assay results yields better clinical outcome than current clinical decision making without molecular assay results. It is equally important to realize that pharmacogenomic search generates a large molecular database of human breast cancer. These data may have an important impact on the future drug and biomarker discovery process independent of the success or failure of the pharmacogenomic diagnostic tests that are currently being evaluated. The current paradigm of therapeutic target discovery starts with experiments in cancer models. Molecular pathways that regulate cancer growth and survival are identified in vitro, inhibitors of key pathway components are then developed and, when appropriate, taken to the clinic for further human testing. High throughput molecular analytic techniques may enable a paradigm shift in this discovery process. The initial target discovery can now be made using human data through the analysis of gene-expression profiles. Promising markers may be taken to the laboratory to gain mechanistic insight into their function and, if appropriate, may be brought back to the clinic as novel diagnostic markers or new drugs. This new discovery process is appealing, because it is more likely to lead to targets that are relevant for human cancer, as opposed to experimental models. Also, not only the target is identified, but a method of defining the patient population for therapy can be developed simultaneously. This process is also relatively unbiased by existing knowledge; therefore, unexpected new discoveries can be made. Acknowledgments Supported by grants from the NCI (RO1-CA106290), the Breast Cancer Research Foundation, the Gilder Foundation, and the Dee Simmons Fund to L.P. C.M. is supported by Fondation de France and the Federation Nationale des Centres de lutte Contre le Cancer, Paris. Disclosure of Potential Conflicts of Interest L.P. and W.F.S. are founders and shareholders of Nuvera Biosciences, Inc. References 1 D Eredita G, Giardina C, Martellotta M et al. Prognostic factors in breast cancer: the predictive value of the Nottingham Prognostic Index in patients with a long-term follow-up that were treated in a single institution. Eur J Cancer 2001;37: Olivotto IA, Bajdik CD, Ravdin PM et al. Population-based validation of the prognostic model ADJUVANT! for early breast cancer. J Clin Oncol 2005;23: Rouzier R, Pusztai L, Delaloge S et al. Nomograms to predict pathologic complete response and metastasis-free survival after preoperative chemotherapy for breast cancer. J Clin Oncol 2005;23: Bast RC Jr, Ravdin P, Hayes DF et al Update of recommendations for the use of tumor markers in breast and colorectal cancer: clinical practice guidelines of the American Society of Clinical Oncology. J Clin Oncol 2001;19: The Oncologist

10 Pusztai, Mazouni, Anderson et al Bonneterre J, Thürlimann B, Robertson JFR et al. Anastrozole versus tamoxifen as first-line therapy for advanced breast cancer in 668 postmenopausal women: results of the Tamoxifen or Arimidex Randomized Group Efficacy and Tolerability study. J Clin Oncol 2000;18: Mouridsen H, Gershanovich M, Sun Y et al. Superior efficacy of letrozole versus tamoxifen as first-line therapy for postmenopausal women with advanced breast cancer: results of a phase III study of the International Letrozole Breast Cancer Group. J Clin Oncol 2001;19: Rhodes A, Jasani B, Barnes DM et. Reliability of immunohistochemical demonstration of estrogen receptors in routine practice: interlaboratory variance in the sensitivity of detection and evaluation of scoring systems. J Clin Pathol 2000;53: Rüdiger T, Höfler H, Kreipe HH et al. Quality assurance in immunohistochemistry: results of an interlaboratory trial involving 172 pathologists. Am J Surg Pathol 2002;26: Goldstein NS, Ferkowicz M, Odish E et al. Minimum formalin fixation time for consistent estrogen receptor immunohistochemical staining of invasive breast carcinoma. Am J Clin Pathol 2003;120: Rhodes A. Quality assurance in immunohistochemistry [comment]. Am J Surg Pathol 2003;27: ; author reply Roche PC, Suman VJ, Jenkins RB, et al. Concordance between local and central laboratory HER2 testing in the breast intergroup trial N9831. J Natl Cancer Inst 2002;94: Perez EA, Suman VJ, Davidson NE et al. HER2 testing by local central and reference laboratories in the NCCTG N9831 intergroup adjuvant trial. Proc Am Soc Clin Oncol 2004;22: Harvey JM, Clark GM, Osborne CK et al. Estrogen receptor status by immunohistochemistry is superior to the ligand-binding assay for predicting response to adjuvant endocrine therapy in breast cancer. J Clin Oncol 1999;17: Lacroix M, Querton G, Hennebert P et al. Estrogen receptor analysis in primary breast tumors by ligand-binding assay, immunocytochemical assay, and northern blot: a comparison. Breast Cancer Res Treat 2001;67: 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 2004;351: Habel LA, Quesenberry CP, Jacobs M et al. Gene expression and breast cancer mortality in Northern California Kaiser Permanente patients: a large population-based case control study. Proc Am Soc Clin Oncol 2005;24:603a. 17 Paik S, Tang G, Shak S et al. Gene expression and benefit of chemotherapy in women with node-negative, estrogen receptor-positive breast cancer. J Clin Oncol 24: /JCO Jansen MP, Foekens JA, van Staveren IL et al. Molecular classification of tamoxifen-resistant breast carcinomas by gene expression profiling. J Clin Oncol 2005;23: Symmans WF, Sotirion C, Anderson SK et al. Measurements of estrogen receptor and reporter genes from micro arrays determine receptor status and time to recurrence following adjuvant tamoxifen therapy. Breast Cancer Res Treat 2005;94(supp 1):308a. 20 Hess KR, Anderson K, Symmans W, et al. Pharmacogenomic predictor of sensitivity to preoperative paclitaxel and 5-fluorouracil, doxorubicin, cyclophosphamide chemotherapy in breast cancer. J Clin Oncol 2006 (in press). 21 Anderson K, Hess KR, Kapoor M et al. Reproducibility of gene expression signature-based predictions in replicate experiments. Clin Cancer Res 2006;12: Perou CM, Sorlie T, Eisen MB et al. Molecular portraits of human breast tumours. Nature 2000;406: Pusztai L, Ayers M, Stec J et al. Gene expression profiles obtained from fine-needle aspirations of breast cancer reliably identify routine prognostic markers and reveal large-scale molecular differences between estrogen-negative and estrogen-positive tumors. Clin Cancer Res 2003;9: Sorlie T, Tibshirani R, Parker J et al. Repeated observation of breast tumor sub types in independent gene expression data sets. Proc Natl Acad Sci U S A 2003;100: Sotiriou C, Neo SY, McShane LM et al. Breast cancer classification and prognosis based on gene expression profiles from a population-based study. Proc Natl Acad Sci U S A 2003;100: Sorlie T, Perou CM, Tibshirani R et al. Gene expression patterns of breast carcinomas distinguish tumor subclasses with clinical implications. Proc Natl Acad Sci U S A 2001;98: Rouzier R, Perou CM, Symmans WF et al. Breast cancer molecular subtypes respond differently to preoperative chemotherapy. Clin Cancer Res 2005;11: Hu Z, Fan C, Oh DS et al. The molecular portraits of breast tumors are conserved across microarray platforms. BMC Genomics 2006;7: McShane LM, Radmacher MD, Freidlin B et al. Methods for assessing reproducibility of clustering patterns observed in analyses of microarray data. Bioinformatics 2002;18: Milligan GW, Cooper MC. An examination of procedures for determining the number of clusters in a data set. Psychometrika 1985;50: Tibshirani R, Walther G, Hastie T. Estimating the number of clusters in a data set via the gap statistic. J R Stat Soc B 2001;63: van t Veer LJ, Dai H, van de Vijver et al. Gene expression profiling predicts clinical outcome of breast cancer. Nature 2001;415: 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 2002;347: Piccart MJ, Loi S, van t Veer L et al. On behalf of the TRANSBIG Consortium. Multi-center external validation study of the Amsterdam 70-gene prognostic signature in node negative untreated breast cancer: are the results still outperforming the clinical-pathological criteria? Presented at the 2004 San Antonio Breast Cancer Symposium (Late breaking session), San Antonio, Texas, December 8 11, Wang Y, Klijn JG, Zhang Y et al. Gene-expression profiles to predict distant metastasis of lymph-node-negative primary breast cancer. Lancet 2005;365: Foekens JA, Atkins D, Zhang Y et al. Multicenter validation of a gene expression-based prognostic signature in lymph node-negative primary breast cancer. J Clin Oncol 2006;24: 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 2004;22: Pounds S, Morris SW. Estimating the occurrence of false positives and false negatives in microarray studies by approximating and partitioning the empirical distribution of p-values. Bioinformatics 2003;19: Mootha VK, Lindgren CM, Eriksson KF et al. PGC-1alpha-responsive genes involved in oxidative phosphorylation are coordinately downregulated in human diabetes. Nature Genet 2003;34: Subramanian A, Tamayo P, Mootha VK et al. Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles. Proc Natl Acad Sci U S A 2005;102: