The rapid advancement of high-information content technologies has resulted in the completion of numerous molecular profiling studies of breast cancers. Global gene expression profiling, massively parallel sequencing (MPS), array-comparative genomic hybridization (aCGH), and reverse phase protein arrays (RPPA) have allowed scientists to profile RNA, DNA, and proteins in hundreds of human breast tumor tissues in a speedy fashion. To fully decipher the mountains of data generated from these ‘omics’ approaches, and to translate the findings into the clinical setting is challenging, but, undoubtedly, is also one of the highest research priorities for the next few years. In this chapter, we will review how genomics have informed our understanding of the heterogeneity of breast cancer and how it is currently being used for prognostication and therapeutic decision-making. The promise of gene expression patterns, when possibly coupled with somatic mutational profiles, is the near future when we will be able to use the detailed tumor-specific, and patient-specific, information as a means to personalize therapy for breast cancer patients.

Breast cancer is a known heterogeneous disease comprised of a growing number of recognized biologic subtypes. Clinicians and researchers have noted the variations in risk factors, response to therapy, and clinical behavior according to hormone receptor status (i.e., Estrogen Receptors [ER] and Progesterone Receptors [PR]) for several decades. More recent data has also implicated HER-positive breast cancers as possessing unique characteristics, such as responsiveness to anthracyclines (1). Traditional single marker approaches to biomarker identification is limited by the fact that seldom is one gene/protein responsible for the entire action of a cellular pathway. Even more importantly, single marker studies do not address the important relationships among and within different pathways, which are increasingly becoming needed to predict tumor behavior and response to therapy.

As mentioned earlier, MPS is a new and powerful tool for the study of human cancers. The first commercially available MPS platform was the 454 technology by Roche Applied Sciences (2). As of today, other platforms available in the market include the HiSeq and MiSeq Systems by Illumina, Ion Torrent PGM and Proton by Life Technologies, and PacBio RS by Pacific Biosciences. Each technology uses a proprietary approach to sequence molecules of DNA; however, all result in the generation of tens of thousands, to even tens of millions, of sequence ‘reads,’ which are then used to reconstruct the genomic DNA sequence (or mRNA sequence) of a gene or genome. Although the clinical integration of targeted sequencing assays may still be a few years away, in 2012, eight published landmark studies all applied MPS and/or other DNA-based ‘omic’ technologies to create a comprehensive catalog of somatic mutational events that are driving breast cancer pathogenesis (3, 4, 5, 6, 7, 8, 9 and 10). This new MPS data, and the older gene expression array data, provide the genetic framework for personalized medicine as follows.

DNA microarrays have been used as a means to better understand tumor biology and to predict outcomes and response to therapy. Gene expression microarrays measure the level
of expression of a particular gene by quantitatively determining the level of mRNA transcripts, which can be initially compared to a reference sample (as in the case of two-color arrays like those produced by Agilent), or directly compared to other tumors or normal tissues (as in the case of one-color arrays like those produced by Affymetrix). The arrays themselves currently include essentially all human genes that are encompassed in thousands to millions of features/probes that are either oligonucleotides, or PCR amplified cDNA inserts. These sequences represent known, unknown but validated, and hypothetical genes, and all approximately 37,000 genes in the human genome can be included (˜25,000 protein coding genes, ˜12,000 non-coding RNAs). A tumor is processed for RNA, which is used to generate either complementary DNA or RNA, labeled with a fluorescent probe. These fluorescent probes are then either directly applied to a gene array alone (one-color arrays), or combined with a second fluorescently labeled reference sample and then both are applied to an array (two-color array, in which, by convention, the color ‘red’ is used for the sample of interest and the color ‘green’ for the common reference sample). The remainder of the assay is basically a Southern blot with nucleic acid hybridization reactions occurring and binding, and with the intensity(s) of the nucleic acids that hybridize to the individual gene probes reflecting the relative amounts of tumor mRNA. In the case of a two-color microarray the ‘green’ signal predominance reflects low expression and ‘red’ predominance reflects high expression of that gene in the tumor relative to the reference (Fig. 29-1). Thus, in a twocolor array it is not the value of the tumor versus the reference that is of greatest interest, but the ratio of tumor/reference for each gene that is used as a quantitative measure of that gene. This value is then used to compare tumor to tumor, and tumor to normal; in this way, once a two-color microarray gives a tumor/reference value, it is used nearly identically to the onecolor microarray absolute intensity values, and thus, once the user gets past these initial different data processing steps, all downstream analysis steps are similar.

Most breast cancer molecular profiling studies have focused upon gene/mRNA expression. Multiple approaches to analyzing gene expression microarrays exist; however, it is self-evident that the answer one gets from genomic analysis depends upon the question one asks. For example, unsupervised analyses (which use no external guide) can identify whether there are molecularly identifiable tumor subtypes, also called “class discovery,” while supervised analyses (analyzed with a particular clinical endpoint in mind) can identify if there are genes correlated with relapsed patients versus those that did not relapse (prognostic profiles), or related to patients who responded to a particular therapy versus those that did not (predictive profiles), which are also called “class comparisons” (11, 12). It is tempting to assume that the individual genes that are identified using these methods may themselves be causal in creating the subtype, or the clinical characteristic, and sometimes they are; however, in truth, the identified genes themselves may be mere proxies for genetic events or pathway activation and may not themselves be the cause. A third category of analysis, called “class prediction,” first uses a prespecified gene list (typically coming from a supervised analysis) and a given sample set (and classification rule) that is then used to assign a new individual tumor to a particular category, such as a genomic-based class; all genomic tests offered in the clinic should have reached the class prediction stage.

The molecular profiling of breast cancer has provided important information for three major questions: (1) Are there subtypes of breast cancer based on biological differences? (2) Are there gene expression profiles that can distinguish poor from good prognosis patients, thus allowing us to make better-informed decisions regarding adjuvant therapy? (3) Are there gene expression profiles that can predict which tumor will respond to a specific therapy? These questions, and others, will be addressed in the following sections.

In 2000, Perou and colleagues used a semi-supervised approach to identify naturally occurring breast cancer subtypes in a population of 40 patients with locally advanced disease treated with neoadjuvant chemotherapy (13). They identified 496 genes termed the “intrinsic gene set” that showed little variance within repeated tumor sample, but high variance across different tumors, and then used this gene set for potential subtype discovery. Among these breast cancers, they found that the patterns of expression of these genes segregated the tumors into four subtypes, and in Sorlie et al. (2001), they identified a fifth possible subtype (14). The five ‘intrinsic’ subtypes are so-called because the gene list that defines them reflects intrinsic properties of breast cancers rather than being contributed by other cell types or augmented by drugs. These subtypes have been consistently identified in independent datasets using multiple different technologies (15, 16, 17, 18, 19, 20 and 21), are conserved across ethnic groups, and are present in preneoplasia (21, 22). Reassuringly, the intrinsic subtypes are segregated by expression of hormone receptors and the genes they regulate (and actually include ER, PR, and HER2), supporting earlier epidemiologic and biomarker studies suggesting that ER positive and ER negative breast cancer are different. At least two hormone receptor positive subtypes were identified and are called “Luminal A” and “Luminal B.” Conversely, there are several subtypes characterized by low expression of hormone receptors, one of which is called the “HER2-enriched” subtype (HER2-E) and another called the “Basallike” subtype (Fig. 29-2A). The fifth subtype, the normal-like, is less clearly a subtype rather than a likely technical artifact possibly caused by too much normal contaminating tissue. A new possible subtype, named Claudin-low, has been recently identified, which is characterized by low to absent expression of cell adhesion genes including Claudin 3, 4, 7, and E-cadherin (23). Although the intrinsic subtypes were identified without any knowledge of outcomes, these subtypes have strong prognostic implications (Fig. 29-2B); in particular, patients with Basal-like, Claudin-low, HER2-E, or Luminal B tumors demonstrate a significantly worse outcome compared to patients with Luminal A tumors in datasets from patients treated with no systemic adjuvant therapy, and in patient sets treated more heterogeneously including adjuvant and neoadjuvant chemotherapy (15, 16, 17, 18 and 19, 24, 25).

A critical aspect of biomarker biology is validation and the intrinsic subtypes have been validated through multiple common findings including similar distributions on many independent datasets, and similar overall risks/prognoses as well (14, 17, 19, 26). Because the clustering methodology for the initial identification of the intrinsic subtypes is suboptimal for everyday clinical classifications, the development of a robust subtyping method for individual patient samples has been an area of active research. One of the promising approaches for reproducible subtype classifications is based upon identifying a subtypes mean expression profile, called a centroid (17, 18, 27). Hu and colleagues developed the Single Sample Predictor (SSP) tool to serve as a first generation, unchanging classifier for individual patient samples; the SSP compares the gene expression profile of an unknown sample to a prototypical profile of each intrinsic subtype and classifies the unknown sample according to the profile it most closely matches. Thus, one sample or hundreds of samples can be objectively classified in a reproducible fashion.

Fresh frozen (FF) tumor samples are usually preferred for microarray experiments, therefore limiting the practicability of these approaches for correlative sciences on clinical trials when often only formalin-fixed paraffin embedded (FFPE) tissues are available. However, Mullins and colleagues demonstrated that there was 94% concordance (33 of 35 matched FF-FFPE pairs) when comparing the subtype assignments by centroid-based algorithm from FF tissue by microarray versus FFPE tissue assayed by real-time quantitative reverse transcription-PCR (qRT-PCR) (28). In 2009, Parker and colleagues developed an improved, opensource intrinsic subtype classifier based on a minimal list of 50 genes, commonly known as the PAM50 (29). In this study, 122 matched frozen-FFPE tumor pairs were subjected to both microarray and qRT-PCR analysis, and a centroidbased predictor was developed that used 50 genes, and which could use either FF RNA (microarray) or FFPE RNA (qRT-PCR). This 50-gene subtype predictor provided significant prognostic value independent of the standard clinical parameters in a test set of 761 patients receiving no systemic therapy, and significantly predicted the pathologic complete response (pCR) in a test of 133 patients treated with neoadjuvant taxane and anthracycline regimens (29).

In the same study, the authors also developed several risk models based upon a Cox Modeling approach, which uses the genomic-determined subtype data and a standard clinical variable (tumor size). A Risk of Relapse (ROR) score could be assigned to each test case (that is, a patient sample) using a Cox model based upon a tumors “distance” to the subtype centroids alone (ROR-S), or using subtype correlation along with tumor size (ROR-T) (29). The most current standard for molecular intrinsic subtyping is the PAM50 50-gene test, which includes the above subtypes except the Claudin-low, and is undergoing extensive clinical validation (30, 31, 32, 33 and 34).

Other approaches to defining the intrinsic subtypes have been attempted using non-microarray-based methods, most often using immunohistochemistry (IHC) (35, 36). However, the accuracy using an IHC-based approach is not as great as the multi-gene expression assays, mostly because IHC assessments can be subjective, and suffer from inter-observer and inter-laboratory variations. Despite the limitations of IHC, performing tumor subtyping via IHC is still valuable and has been adopted as a means of classification by the St. Gallen’s consensus conference (37); specifically, the indication is that Luminal A patients as defined by this IHC-based definition (i.e., ER+ and/or PR+, HER2-normal, Ki-67 less than 14%), may not be recommended to receive adjuvant chemotherapy given their overall general good prognosis.