Overview: This report reviews the findings published by Ma et al., in two papers entitled: Gene Expression Profiles of Human Breast Cancer Progression, and A two-gene expression ratio predicts clinical outcome in breast cancer patients treated with tamoxifen; PNAS USA 100(10) 2003 and Cancer Cell 5(Jun) 2004, respectively. Ma et al., combined the precision of laser capture microdissection and the high throughput capacity of microarrays, to characterize gene expression signatures associated with the various pathologically defined stages of breast cancer to identify changes in gene expression that are potentially responsible for breast cancer progression. The second study uncovered a simple two-gene expression ratio that predicts clinical outcome of patients having estrogen receptor (ER) positive breast cancers treated with Tamoxifen. Significantly, this novel two gene expression-ratio is a far more powerful predictor of clinical outcome than any existing clinical or molecular marker.

Introduction: A principal goal of oncogenomics is to understand the molecular changes that underlie tumorigenesis and cancer progression. Tumorigenesis (of the breast), is the phenotypic manifestation of the acquisition and accumulation of genotypic alterations, and is thought to be reflected by pathologically-defined stages of (breast) cancer progression(1). Most breast cancers arise in a structure called the Terminal Duct Lobular Unit (TDLU) and are derived specifically from luminal epithelial cells, which form the lining of the TDLU (16). To summarize the authors' example, tumorigenesis may be initiated when a phenotypically normal luminal epithelial cell enters a premalignant stage of growth called atypical intraductal hyperplasia (ADH). Participant ADH cells then progress into a proliferative phase giving rise to the malignant, but preinvasive lesion, ductal carcinoma in situ (DCIS). The growing tumor may then degrade basement membranes and become invasive, moving into adjacent tissues (invasive ductal carcinoma or IDC). The pathologically-defined stages of breast cancer progression are depicted above in a figure adapted from Ma et al., 2003. Interestingly, lesions of the DCIS and IDC stages within single, or among multiple individuals, appear heterogeneous with respect to various cytologic characteristics including mitotic index and differentiation status. Tumor grading schemes have been developed to classify DCIS and IDC tumors using criteria such as 'degree of differentiation'. These subclasses, grades I, II, and III (low, intermediate, and high), are associated with well, moderate, and poorly differentiated tumors, respectively (4,9).

Grading systems are used by clinicians when making patient-care decisions because of correlations observed between clinical outcome and tumor grade. Patients with high-grade tumors usually have a poor prognosis whereas those with low-grade tumors demonstrate more favorable outcomes (4,14). Additional prognostic indicators, such as the tumors' expression of the estrogen and progesterone receptors (ER and PR), or the epidermal growth factor receptor Her2/Neu, sometimes aid clinicians when choosing an appropriate treatment for their patient. For example, breast tumors that express high levels of Her2/Neu protein are often associated with poor clinical outcome and mandate aggressive combinatorial chemotherapies (10). In contrast, the expression of ER+ and/or PR+ on breast tumors indicates that the tumor may be hormone-dependent, that is, it requires a functional ER/PR signaling pathway to maintain the malignant phenotype (2). Thus, patients having ER+/PR+ breast cancers usually respond well to endocrine therapy such as Tamoxifen, and may not require additional chemotherapy to kill the cancer cells (7). Tamoxifen is an estrogen analog that acts by competitive inhibition of estrogen binding and signaling through the ER pathway. Treatment with Tamoxifen confers a 40%-50% reduction in the annual risk of tumor recurrence (5). This translates into a 5%-10% improvement in 10-year survival for patients (5). Expression of ER and/or PR are currently the best predictors of tamoxifen response (2,5). However, many ER+ breast tumors fail to respond to Tamoxifen or quickly develop resistance to treatment (3,13). For such cases, aromatase inhibitors, or agents such as gefitinib that target other pathways, might be effective in improving clinical outcome (6,10). Unfortunately, there are no sound markers that identify patients who would be better candidates for a given therapy. Thus, the identification of the transcriptional programs associated with the initiation and progression of breast cancer and, those indicative of specific therapeutic responses, should greatly enhance the management of breast cancer and likely improve clinical outcome.

The first microarray experiments designed to characterize the gene expression programs of breast cancer were performed by Sorlie. This landmark study confirmed earlier observations that, breast cancers expressing the estrogen receptor were more likely to be associated with good prognosis than ER- breast cancers and that tumors are likely clonal because the gene expression profile of one tumor is usually more similar to the profile of an additional sample from the same, rather than a different, tumor. Interestingly, the gene expression signatures obtained from various breast tumors were characterized by high expression levels of lymphocyte, adipocyte, and stromal cell -specific genes. They determined that these signatures were 'molecular portraits' of each tumor because they captured all of the unique elements that contributed to the tumors' environment. However, due to the absence of patient-matched normal tissue controls and the cellular complexity of the biopsy specimens interrogated, the experiments precluded drawing any definite conclusions regarding genes expressed by the cancer cells themselves, or what their level of expression was, relative to the patient's normal cells.

The collaborative studies described below, led by Mark Erlander Ph.D. and Chief Scientific Officer of Arcturus Engineering, and Dennis C. Sgroi M.D. of Massachusetts General Hospital, sought to characterize gene expression profiles associated with the different tumor grades and stages of breast cancer progression that are frequently encountered by clinical oncologists. They hypothesized that the histologic and pathologic differences between benevolent and aggressive tumors might also be reflected by differences in gene expression. They reasoned that such differences could potentially be used to classify tumors that lack the standard histo-pathologic features currently utilized for classification. To properly test these hypotheses, Ma et al. combined Laser Capture Microdissection (LCM) and Linear RNA amplification technology in order to examine samples that were free from confounding artifacts such as infiltrating lymphocytes and stromal tissue.

Results:(Ma et al., 2003)
Validation of LCM and Linear RNA Amplification Utility in Gene Expression Profiling Applications. A major goal of this study was to determine the gene expression signatures of each of the pathologic stages of breast cancer progression, and to identify expression differences associated with the transition from pre-invasive to invasive lesions. Previous microarray studies used whole clinical biopsies to examine differences in gene expression (15). Such an approach was unsuitable for this study because the majority of cells within clinical biopsies are largely non-tumor, that is, the cells comprising the cancer itself are limiting. To circumvent difficulties isolating tumor cells imposed by the cellular complexity of biopsies and to obtain sufficient RNA to perform microarray experiments, Ma et al. combined LCM (Fig.1) and T7-based linear RNA amplification. Earlier attempts to amplify minute quantities of RNA were biased, rather than uniform in amplification, raising valid concerns about the fidelity of RNA amplification in maintaining transcript levels representative of the original unamplified sample. To address this concern, LCM was performed on adjacent tissue sections of patients from whom abundant tissue was available (215-Normal; 89, 178, 179-DCIS; 97, 169, 170-IDC). For each case, LCM was first used to procure 40,000 normal or tumor cells from which total RNA was isolated and converted to ds-cDNA directly with no amplification. LCM was then applied to adjacent sections from the same cases to harvest ~2500 normal or tumor cells from which total RNA was isolated and amplified using the RiboAmp kit (Arcturus). Taqman real-time quantitative PCR was used to assess and compare the level of expression of 5 genes, in each of the amplified or unamplified DCIS and IDC cases relative to the amplified or unamplified normal sample. Linear regression analysis of the gene expression data was used to determine the degree of concordance among the amplified and unamplified samples. An R² value of 0.96 illustrated that T7-based Linear RNA amplification preserves the differential expression of genes between samples. This finding indicates that combining the use of LCM and linear RNA amplification is a dependable strategy for gene expression profiling applications (Fig.5 supplementary data).
Gene Expression Profiles of Breast Tumor Stages.To examine the gene expression profiles characteristic of the sequential disease stages (ADH, DCIS, and IDC) of breast cancer, Ma et al. used LCM to capture cells constituting only the tumor lesion and adjacent (patient-matched) phenotypically normal epithelium (PNE). Tissue from 3 elective mammoplasty-reduction patients served as disease-free normal controls. Each component (PNE, ADH, DCIS, and/or IDC) was microdissected in triplicate for each patient. RNA from each individual sample was isolated and amplified independently (PicoPure and RiboAmp kits, Arcturus), and was then competitively hybridized to cDNA microarrays in duplicate against human Universal Reference RNA (stratagene) that had been amplified in parallel. Cluster analysis of all samples relative to the universal reference RNA demonstrates that all of the PNE samples from breast cancer patients are most similar to the mammoplasty-reduction patients (see below). This suggests that patient-matched PNE is an excellent control, representing the patient's baseline gene expression profile, for assessing differences in gene expression associated with breast cancer progression. The cluster analysis indicates that the majority of differences in relative gene expression levels occur early during the ADH stage of disease progression.
Next they narrowed their focus to ~1900 genes out of the 12000 by selecting those with discriminant function coefficients >0.5. The data for that gene set is presented as the expression ratio of disease-state relative to patient-matched normal. Hierarchical clustering of the disease-state samples identified 2 gene clusters, one of which represents global decreases in transcript abundance, and one that demonstrates an increase in transcript abundance (Fig.2). Upon hierarchical clustering analysis, samples of different disease-states from the same patient cluster more closely than to samples of similar disease-states from other patients. Microarray data were verified by performing RTQ-PCR on a number of differentially-expressed genes for all patients. Thus it appears, that the transcriptional changes characteristic of the invasive stage of breast cancer progression, are already present in the pre-malignant stage (ADH).
Gene Expression Profiles of Breast Tumor Grade. To determine if breast tumors of different grade also had distinct gene expression patterns, a discriminant analysis of Low grade and High grade tumors was performed to identify the top 100 genes characteristic of those grades. The expression for each gene was determined as the ratio of disease-state to patient-matched normal, and the data were subjected to clustering analyses for comparison. The 2-dimensional clustering identified three main clusters of genes(Fig.3). The Grade I cluster consists of genes increased in grade I tumors whereas the Grade III signature identifies genes whose expression is increased in high grade tumors. The third cluster represents genes that are decreased in abundance across all of the samples. An interesting observation is that some moderate grade tumors cluster more closely with low grade tumors while others exhibit expression signatures characteristic of high grade tumors. These data are significant because they demonstrate that moderate grade tumors (grade II) appear to represent a transition-state from low-to-high grade disease.
The Relationship of the DCIS-IDC Stage Transition to Tumor Grades.Finally, the relative expression of genes between patient-matched IDC and DCIS was determined for the ~1900 gene set and subjected to 2-dimensional clustering. This analysis identified a small subset of genes (29), that were consistently increased in the patient's IDC component compared to their DCIS. These differences were even more apparent in grade III tumors when compared to tumors of lower grade (Fig.4). The data demonstrate that the expression of a significant number of genes is increased in grade III DCIS compared to grade I DCIS and that the expression of these genes is augmented in the IDC component of the tumor.

Results:(Ma et al., 2004)
Identification of differentially-expressed genes. Ma et al., performed microarray analyses on material derived from whole frozen tissue sections from 60 patients with ER+ breast tumors (Cohort 1) to identify possible gene expression signatures that were predictive of clinical outcome in response to treatment with tamoxifen. The resulting data were filtered according to variance such that genes in the 75th percentile were investigated further. These genes were subjected to a t-test that compared patients from the recurrence group with those from the nonrecurrence group. The t-test resulted in 19 genes (Fig.1) that had a significance cutoff of p=.001. The analysis was repeated, this time using material derived from laser capture microdissected frozen tissue sections from the same 60 patients. LCM was used to avoid confounding expression artifacts contributed by non-tumor cells. This analysis identified 9 genes with a p value of less than p=0.001. Three of the nine genes identified in the LCM analysis were also among the 19 genes identified when whole tissue sections were used. HOXB13 encodes a homeobox transcription factor, and demonstrated increased transcript levels in the tamoxifen recurrence group. IL17BR encodes the interleukin-17B receptor and was overexpressed in the tamoxifen nonrecurrence group. Finally, AI240933, an expressed sequence tag (EST), was also overexpressed in the tamoxifen nonrecurrence group. Ma et al., then determined the ability of HOXB13, IL17BR, and AI240933 to function as prognostic factors predicting clinical outcome of patients treated with tamoxifen. Receiver Operating Characteristic (ROC) analysis was used to assess the specificity of these markers. The Area Under the Curve (AUC) values obtained for each marker were significantly greater than the value 0.5, the value for randomly predicting clinical outcome as calculated by the null model (Table 2). To compare the effectiveness of HOXB13, IL17BR, and AI240933 to predictors of tamoxifen response that are currently the standard of care, ER and PR expression data was obtained from the arrays and subjected to the ROC analysis. ERBB2 expression data was also retrieved and analyzed because it is inversely correlated with tamoxifen response (10). AUC values were only significant for PR and ERBB2, but the analysis did assert previous correlations with clinical outcome (Table 2; 2,10). These findings demonstrate that AI240933, IL17BR, and HOXB13 all outperform the current standard of care markers for predicting tamoxifen response.
HOXB13:IL17BR expression ratio is a robust composite predictor of outcome. Having demonstrated that HOXB13, IL17BR, and AI240933 each were superior predictors of outcome relative to current predictors, Ma et al., then tested whether or not using a combination of the three genes would be a more powerful predictor than using any one alone. First, HOXB13 and IL17BR were selected for composite analysis because their increased expression is associated with different outcomes. T test and ROC analysis both confirm that the ratio of HOXB13/IL17BR expression is a superior predictor of tamoxifen response when compared to either gene alone (Table 3). The HOXB13/IL17BR ratio attained AUC values of 0.81 and 0.84 in the whole-tissue and LCM-tissue datasets, respectively, compared to 0.79, the highest AUC value obtained for either HOXB13 or IL17BR alone. Apparently, combining AI240933 with either HOXB13, or both HOXB13 and IL17BR, failed to confer additional power as a predictor of tamoxifen response. Next, Ma et al., compared the predictive power of the HOXB13/IL17BR expression ratio to traditional prognostic indicators in breast cancer like tumor size, grade, node status, and patient age (8). Because the initial experiments were designed such that the patients comprising the recurrence and nonrecurrence groups were closely matched in terms of tumor grade, size, and lymph node status, the only factor reaching significance (barely) was tumor size. Earlier, they had determined that only PR and ERBB2 expression were significant predictors of tamoxifen outcome (Table 2). Ma et al., performed a comparison of logistic regression models representing HOXB13/IL17BR expression alone or in combination with the expression of PR and ERBB2, or tumor size. They determined that the HOXB13/IL17BR expression ratio was the sole variable attaining significance in the multivariate model (p=0,002, Table 4). This finding suggests that the HOXB13/IL17BR expression ratio is highly predictive of tamoxifen treatment outcome and, can be used as a predictor independent of other common prognostic factors.
Independent validation of HOXB13/IL17BR expression ratio to predict tamoxifen treatment outcome. The identification of a new marker that predicts clinical response of early stage ER+ breast cancer patients to tamoxifen treatment, with more accuracy than all other markers currently available, may prove useful to clinicians. However, to have any impact on patient care, the process by which this novel predictor is assessed, needs to be transformed into an assay that can be performed in the average clinical setting. Because the signature predicting outcome was refined to the expression ratio of only two genes, Ma et al., reasoned that simpler assays for examining gene expression might be equally successful as microarrays. Thus, Ma performed Taqman RTQ-PCR to examine HOXB13 and IL17BR expression in tissue sections derived from 59 of the original patients (Fig 2A). The RTQ-PCR results correlated well with those obtained from the microarray experiments (Fig 2B and 2C). The Pearson correlation coefficient for IL17BR expression between the RTQ-PCR and microarray data was r=0.93, while HOXB13 received a r=0.83. Importantly, the HOXB13/IL17BR expression ratios calculated from the RTQ-PCR data also demonstrated a high degree of concordance with the microarray data (r=0.83) and tamoxifen treatment outcome (t test p=5.8e-06). To validate the HOXB13/IL17BR expression ratio as a predictor of tamoxifen treatment outcome, Ma went on to perform the RTQ-PCR analysis on an independent group of patients (Cohort 2). Of note, these validation RTQ-PCR analyses were performed on material derived from tissue that was formalin-fixed and paraffin-embedded (FFPE), demonstrating the efficacy of the assay for applications using standard clinical specimens. In agreement with the frozen cohort data, the HOXB13/IL17BR expression ratio proved to be significantly associated with outcome in the FFPE cohort (p=0.024, Fig 2D). To assess the predictive power of the HOXB13/IL17BR ratio, a logistic regression model based on the RTQ-PCR data of cohort 1 was applied to the RTQ-PCR data set of cohort 2. The expression ratios obtained from the FFPE specimens correctly predicted the clinical outcome for 16 of the 20 patients treated with tamoxifen in cohort 2 (Fig 2D). Lastly, Ma et al., use Kaplan-Meier survival curves to compare the lengths of disease-free survival (Fig 2E and 2F) of patients segregated by the HOXB13/IL17BR expression ratio cutoff point (Fig 2C and 2D) to demonstrate the power of their marker for predicting clinical outcome. Together, these data illustrate that the HOXB13/IL17BR expression ratio, as a predictor of tamoxifen treatment outcome, was similarly effective when obtained from frozen or FFPE tissues derived from independent patient cohorts.

Summary and Concluding Remarks: The findings reported by Ma et al. (2003 and 2004), and summarized here, illustrate how technological advances in molecular biology are enabling scientists to understand (or at least explain), some of the apparent discrepencies observed between tumor phenotype and behavior, that have confounded clinical oncologists for years. In an effort to make pathological diagnostics more consistent, tumor grading systems were developed that utilize specific cytologic and pathologic features of the tissue as classification criteria. Indeed, such prognostic indicators have been instructive for prescribing different therapeutic strategies for patients with the same stage of cancer due to the correlations observed between well differentiated (grade I) tumors and positive outcome, and poorly differentiated (grade III) tumors and negative outcome. The criteria used by pathologists to stage and grade tumors have proved fairly reliable for those tumors that fall at either ends of the grading spectrum. Unfortunately, a large portion of tumors exhibit features that exclude them from either grade I or III groups making it difficult for clinicians to select the appropriate therapy for these grade II tumors. Why is it that grade II tumors respond differently to specific treatment? The microarray data presented in Ma et al. (11) give us a clue to the reason behind the apparent inconsistent behavior of grade II tumors. Their data identify expression signatures that clearly distinguish grade I and grade III tumors. When grade II tumors are included in the cluster analysis, their distribution is strikingly bipolar. Half of the intermediate grade tumors cluster closely with low grade tumors while the remaining half clusters well with the high grade tumors. The expression signature of grade II tumors suggests that they may reflect a transition state of progression from low to high grade cancer rather than a seperate disease. This expression signature could potentially be used by clinicians to determine the most appropriate therapy for a given grade II tumor and thus increase the likelihood of favorable outcome for patients.

The factors responsible for pre-invasive malignant tumors attaining invasive competency are largely unkown. Ma et al., were able to investigate changes in gene expression among the different stages of cancer progression within single patients because LCM enabled them to harvest tumor cells while excluding non tumor cell types. This specificity gave them a clear picture of the tumors' gene expression profile as each progressed from the premalignant through invasive stages of breast carcinoma. They identified consistent quantitative differences in the expression of a subset of genes when comparing DCIS and IDC components. They also noted that these genes exhibited differences between DCIS of different grades. According to Ma et al., one possible interpretation of these findings suggests that breast cancer stage, and grade progression, are not necessarily independent events in disease evolution. To illustrate their reasoning, they propose a 2-dimensional model of breast cancer progression (adapted from Dr. Sgroi, personal communication;see below). As the authors state, support for such an interpretation is derived from clinical observations that high grade DCIS has a greater propensity for association with invasive cancer than its low grade equivalent. However, the most unexpected finding of their study was that global increases and decreases in gene expression associated with advanced breast cancer disease seem to occur at the earliest, pre-malignant stage of progression (ADH). It is tempting to speculate that the gene expression programs conferring invasive growth are pre-specified and already activated in the preinvasive tumor.

The inability of existing clinical prognostic indicators to accurately predict patients' outcome in response to specific therapy is increasingly problematic given the risks entailed in the trial and error approach of chemotherapeutics. Oncologists know well that despite expressing the markers considered most predictive of response to tamoxifen, some 40% of all the ER+ and/or PR+ early stage breast cancers they treat are unresponsive or become resistant to tamoxifen. For these patients, identification of the HOXB13/IL17BR expression ratio as an independent prognostic indicator of clinical outcome in response to tamoxifen treatment comes none to soon. Ma et al., demonstrate the HOXB13/IL17BR ratio, as a predictor of tamoxifen treatment outcome, is far superior to all other clinical predictive markers. Furthermore, they demonstrate that RTQ-PCR is equally capable of measuring the HOXB13/IL17BR expression ratio without losing its predictive power. Ma et al. then illustrated the clinical utility of their diagnostic assay by successfully applying it to both frozen and FFPE clinical specimens. If the HOXB13/IL17BR ratio prevails in future studies, it has potential to identify those ER+ breast cancer patients who risk tumor recurrence in the context of tamoxifen treatment. As such, the HOXB13/IL17BR signature could have potential utility for identifying patients that might derive greater benefit from a more aggressive or combinatorial therapeutic approach. Together, the experiments performed by Ma et al represent important steps towards understanding the molecular basis of human breast cancer progression. However, their most significant contribution to oncogenomics lies within the identification of a new and independent molecular signature that is unmatched when used to predict an ER+ breast cancer patient's therapeutic response to tamoxifen treatment. The proof-of-principle that gene expression can reliably predict clinical outcome is vital to the advancement and improvement of disease management and patient care.

References:
1. Allred et al. (2001), Endocrinol. Relat. Cancer 8, 47-61
2. Bardou et al. (2003), J. Clin. Oncol. 21, 1973-9
3. Clarke et al. (2003), Oncogene 22, 7316-7339
4. Dalton et al. (2000), Mod. Pathol. 13, 730-5
5. EBCTCG (2001), Cochrane Database Syst. Rev. 1, CD000486
6. Ellis et al. (2001), J. Clin. Oncol. 19, 3808-16
7. Fernandez et al. (1983), Clin. Oncol. 9, 245-50
8. Fitzgibbons et al. (2000), Arch. Pathol. Lab. Med. 124, 966-78
9. Holland et al. (1994), Semin. Diagn. Pathol. 11, 167-80
10. Lichtner (2003), Biomed. Pharmacother. 57, 447-51
11. Ma et al. (2003), PNAS USA 100, 5974-5979
12. Ma et al. (2004), Cancer Cell 5, 607-616
13. Nicholson et al. (2003), Breast Cancer Res. Treat. 80 (Suppl 1), S29-S34
14. Page et al. (2001), Am. J. Clin. Oncol. 24, 10-18
15. Sorlie et al. (2001), PNAS USA 98, 10869-10874
16. Wellings et al. (1973), J. Natl. Cancer Inst. 50, 1111-18