Molecular Classification of Breast Carcinoma

Introduction

Historically, breast cancers have been classified broadly based on morphology into ductal and lobular types and characterized by tumor grading. A smaller percentage of tumors are classified as “special subtype carcinomas.” This classification system has provided useful prognostic information. Long-term follow-up studies have shown an excellent prognosis for Nottingham grade I tumors and a poor survival rate for Nottingham grade III tumors. Although extremely useful, the morphological classification does have several drawbacks. First and foremost, there is no difference in disease-free and overall survival between stage/grade-matched ductal and lobular tumors. Second, investigators have used different criteria, mostly arbitrary, to define special subtype tumors. Last but not the least, the most important part of morphological classification, grading, is limited by interobserver variability. All of these factors plus the need to identify new prognostic and predictive factors prompted the efforts on molecular classification of breast carcinoma with the availability and development of molecular profiling assays.

Molecular characterization and classifications of breast carcinomas have wide applications in (1) diagnosis of special tumor types, (2) tumor prognosis, (3) prediction of therapeutic response, (4) identification of novel therapeutic targets, (5) monitoring disease progression, and (6) improving our knowledge of underlying biologies of breast carcinomas which encompass a diverse disease spectrum. This chapter summarizes current understanding and updates on molecular classification of breast carcinomas and molecular prognostic assays.

Intrinsic Gene Set–based Molecular Classification

The breakthrough came in the year 2000 when Perou and colleagues described the breast cancer molecular portraits using cDNA microarrays ( Fig. 20.1 ). Perou et al used a cDNA microarray to study 8,102 human genes on 65 surgical specimens (36 invasive ductal carcinomas, of which 20 were sampled twice and 2 tumors were paired with lymph nodes; 2 lobular carcinomas; 1 case of ductal carcinoma in situ [DCIS]; 1 fibroadenoma; and 3 normal breast tissues). The experimental (tumor) sample was fluorescently labeled with Cy5 and the reference (mRNA pooled from 11 different cell lines) was labeled with another fluorescent nucleotide called Cy3. These were hybridized to the cDNA microarray and the relative abundance of the two transcripts was visualized using a pseudocolored image by a ratio of “red” to “green” fluorescence intensity at each spot. The ratios were log transformed and the data were further analyzed by a hierarchical clustering algorithm that classifies samples based on their overall similarity of gene expression patterns. The authors chose 1,753 of the total 8,102 genes for this purpose as these 1,753 gene transcripts showed a fourfold variation over the median abundance in their sample set in at least three samples. The “molecular portraits” thus created showed similarities and differences among breast tumors related to biological variables such as variations in growth rate, activities of specific signaling pathways, and cellular composition of tumors. However, it was realized that the gene set of 1,753 genes was not optimal to show the intrinsic biological properties of the tumor. A key aspect of this study design was the inclusion of 22 tumors that were sampled twice. Twenty tumors had samples obtained before and after doxorubicin chemotherapy and two tumors were paired with lymph nodes. Fifteen of the 20 paired (before and after) samples had clustered together. These findings implied that each tumor is unique and has a distinct gene expression signature. It also implied that the type and number of nonepithelial cells in carcinoma remain fairly constant and appear not to interfere in expression analysis. Based on these findings, the authors selected an alternative gene subset for testing. This resulted in the selection of 496 genes—an intrinsic gene subset consisting of genes with significantly greater variation in expression between different tumors than between paired samples of the same tumor.

The intrinsic gene set was ideally suited for classification because it consisted of genes whose expression patterns were characteristic of an individual tumor as opposed to those which vary as a function of tissue sampling. In further analyses of up to 115 carcinoma samples, the intrinsic gene set–based molecular classification revealed five distinct classes of breast carcinoma ( Box 20.1 ): luminal A, luminal B (which likely also included the initially described luminal C), ERBB2, basal like, and normal breast like (note that the list of genes varied from 456 to 534 in the initial three studies from 2000 to 2003). The luminal tumors were named as such due to the high expression of genes normally expressed by the luminal epithelium of the breast. These luminal tumors also expressed estrogen receptor (ER) and ER-related genes (LIV1, GATA3, HNF3A, and X-box binding protein 1), with luminal A tumors showing the highest expression of ER. The luminal B and luminal C tumors expressed ER cluster genes at a lower level, but the luminal C tumors also expressed some unique genes (GGH, LAPTMB4, NSEP1, CCNE1) whose coordinated function is unknown. The other three subtypes constitute the ER– group. The ERBB2 tumors, as the name implies, were characterized by the expression of genes in the ERBB2 (or human epidermal growth factor receptor 2 [HER2]) amplicon at 17q22.24. The basal-like tumors were named as such as they expressed genes expressed by the basal cells of stratified epithelium and myoepithelial cells of the breast. Therefore, the basal-like tumors were characterized by the expression of keratin genes KRT5 and KRT17. The normal breast–like group was described to express genes known to be expressed by adipose tissue and other nonepithelial cell types. These tumors were also shown to express basal epithelial genes. However, it is argued that this was likely an artificial category due to poorly sampled tumor tissue. The validity of this classification system was further tested by other investigators with respect to overall and relapse-free survival. The basal-like and ERBB2 tumors showed the worst outcome, luminal A showed the best outcome, and luminal B (including luminal C) was intermediate.

In order to transform the research findings of the intrinsic gene set–based breast cancer classification to a clinically validated assay, Hu and colleagues used publicly available breast cancer gene expression data sets and a novel approach to data fusion. They also described a new intrinsic gene list. This new gene list was created in order to take advantage of the significant advancement in technology in the six years since the first description in the year 2000. The investigators used Agilent oligo microarrays containing 17,000 genes (compared to the 8,000 genes initially evaluated). As mentioned previously, the initial array-based gene expression studies utilized paired samples before and after 16 weeks of chemotherapy, which likely altered the posttherapy sample predominantly in terms of proliferation. Therefore, Hu et al used a 105-tumor training set containing 26 untreated paired samples to derive a new breast tumor intrinsic gene list. These experiments yielded an intrinsic gene set of 1,410 microarray elements representing 1,300 genes including a proliferation signature that was not present in previous breast intrinsic gene sets. The new Intrinsic/University of North Carolina (UNC) gene set showed an overlap of 108 genes with the prior Intrinsic/Stanford gene set of Sorlie et al. The most significant difference was the presence of a large proliferation signature in the new gene set as it was based upon mostly pretreatment paired samples (22/26 were pretreatment pairs). These findings confirmed that proliferation rates of tumors are altered after chemotherapy, but proliferation itself is an intrinsic feature of a tumor’s expression profile.

The 1,300 Intrinsic/UNC gene list derived from the training set was then applied to a combined data set of 315 samples (311 tumors and 4 normal samples) that were taken from three previous experiments (Sorlie 2001 and 2003; Van’t veer et al 2002; Sotiriou 2003). These data sets were combined into one data set first by finding the common genes present in all the data sets and then by using a multivariate analysis tool called distance weighted discrimination . Finally, they identified only 306 of the 1,300 genes in the Intrinsic/UNC gene list to be present in all samples. Despite the loss of more than half the genes in the Intrinsic/UNC gene list, the hierarchical clustering analysis of 315 samples in the combined data set yielded similar classes as described using the Intrinsic/Stanford gene list.

Additional analysis also showed significant association between molecular subtype and standard parameters. The basal-like and HER2-enriched molecular classes contained significantly more tumors graded as grade III. A significant association was also found between ER status and molecular subtype. As expected, no association was found between lymph node status and subtype, and similarly, tumor size and molecular subtype were not strongly correlated. The average age of patients with a luminal A tumor was significantly higher, which is concordant with the well-known fact that most elderly patients develop tumors that are diffusely and strongly positive for ER.

Another significant part of this study was the development of a single sample predictor (SSP) method for assigning a new sample to the aforementioned molecular subtypes. It was necessary to develop the SSP method as hierarchical clustering (a class discovery tool) cannot be applied to any new sample without reanalysis of all the samples. The SSP method required creation of a mean expression profile for each subtype, called the centroid . It is important to note that only samples that were clearly assigned the subtype in hierarchical clustering (i.e., 249 of the 315 samples) were used to create the centroids. The new samples were compared to the centroid of each subtype, and were assigned to a subtype to which they most closely resembled using Spearman correlation. This method was further validated on two additional data sets not previously used (Ma et al 2004; Chang et al 2005). Recurrence-free survival was best for luminal A tumors and worse for luminal B, HER2-enriched, and basal-like tumors in both data sets. The SSP method was also applied to 105 tumors used for the training set. A subset analysis of 48 cases with immunohistochemical (IHC) results revealed an excellent association between clinical hormone receptor (HR) and HER2 determination and subtype assignment using the SSP method.

In a companion publication, the same group of investigators compared microarray-based gene expression profiling with quantitative reverse transcriptase polymerase chain reaction (qRT-PCR) for 53 genes on fresh frozen tissues for classification and risk stratification of invasive breast carcinoma. The 53 genes were selected due to their importance in making “intrinsic” subtype distinctions and their association with cell proliferation. Microarray-based profiling and qRT-PCR were performed on 123 samples containing 117 carcinomas, and the prognostic significance of subtypes was further assessed on publicly available data sets containing 337 samples with long-term follow-up (median 86.7 months). The investigators found 93% concordance between qRT-PCR assay- and microarray-based expression profiling in determining intrinsic subtypes. The intrinsic subtypes identified by either methodology were predictive of outcome. The 14 proliferation genes used in the qRT-PCR analysis were used in creating a single meta-gene which provided prognostic information for patients with luminal subtype tumors. High proliferation in the luminal subtype conferred a 19-fold relative risk of relapse compared with luminal tumors with low proliferation. This study showed that an abbreviated gene set examined by qRT-PCR can recapitulate microarray-based classification of breast carcinoma and can risk-stratify patients using the intrinsic subtype and proliferation.

In order to make it even more applicable to routine pathology specimens, a risk prediction model based on intrinsic molecular subtypes was described 3 years later. This model utilized a 50-gene qRT-PCR assay that could be applied to formalin-fixed paraffin-embedded (FFPE) routine pathology samples. For this analysis, Parker et al analyzed 189 tumors by expression profiling using an expanded intrinsic gene list of 1,906 genes from previous publications. Quantitative qRT-PCR analysis was performed on 122 of the 189 samples for 161 genes that passed the FFPE performance criteria. Finally, 50 genes were selected that had the lowest cross-validation error. The centroids were constructed by using a prediction method described by Tibshirani and colleagues called Prediction Analysis of Micro-array (PAM) because of its reproducibility in subtype classification. The distances were calculated using Spearman’s rank correlation. This method of subtype prediction was named PAM50. This qRT-PCR assay was validated on test sets of 761 cases which showed the prognostic value of molecular subtypes. Using statistical models, a risk of relapse (ROR) score was assigned to each test case using correlation to subtype alone (ROR-S) or using subtype correlation and tumor size (ROR-C). The ROR-C model resulted in improved risk prediction for relapse in untreated patient cohorts compared with either subtypes or clinical markers alone. The ROR-S model predicted neoadjuvant chemotherapy (taxane and anthracycline regimen) efficacy with a negative predictive value for a pathologically complete response (pCR) of 97%. Unlike some prior studies, one prominent observation in this study was the significant discordance between clinical receptor status and molecular subtypes. Of the 626 ER+ tumors analyzed in the microarray test set, 73% were luminal (an ER+ molecular subtype), 11% were HER2-enriched (an ER– molecular subtype), 5% were basal-like (an ER– subtype), and 12% were normal-like (also supposed to be ER–). Conversely, the ER– tumors comprised 11% luminal, 32% HER2-enriched, 50% basal-like, and 7% normal-like. There could have been several valid reasons for these discrepancies, such as tumor heterogeneity with respect to receptor status; very low degree of HR positivity, which by defined/accepted criteria are unfortunately labeled as HR+ tumors; and admixture of normal or nontumoral tissue resulting in dilution of mRNA used for expression analysis, among others. But instead of examining or providing details, the investigators concluded that ER and HER2 status alone are not accurate surrogates for true intrinsic subtype status.

A recent similar publication also concluded that the PAM50 gene expression test for intrinsic biological subtype can be applied to large series of FFPE breast cancers, and gives more prognostic information than clinical factors and IHC using standard cutpoints. Once again the key words were standard cutpoints , which create artificial categories (such as ER+ and ER– tumors based on 1% cutoff). For example, a tumor expressing ER weakly in 1% of tumor cells is closer to an ER– tumor than to a tumor showing strong expression in nearly every cell. Despite this fact, the tumor showing 1% weak expression is considered an ER+ tumor in the above mentioned studies and also for statistical analysis where ER status is considered a categorical variable. This minor deviation then results in a significant discrepancy rate for which details are not provided in the manuscripts or even in supplementary data as semiquantitative results are not always available. ER and HER2 IHC status alone certainly has limitations in predicting molecular subtype, but almost none of the above publications related to molecular subtyping made an attempt to directly compare the prognostic/predictive value of subtypes with the combined power of histological grading, semiquantitative IHC results, and clinical parameters. A bigger question is whether an expensive nonmorphological molecular test is required in lieu of IHC and histological grading in all cases.

There have also been concerns about reproducibility of different platforms in predicting the molecular subtype of an individual sample. Weigelt et al assessed the clinical usefulness of three different SSPs by comparing different methods of breast cancer molecular subtype assignment to ascertain whether each SSP identified molecular subtypes with similar associations with outcome. For this purpose, the investigators analyzed 53 microdissected in-house samples and three cohorts of breast cancer samples in the public domain. The public domain data sets utilized in this study were the NKI-295 ( n = 295), Wang et al ( n = 286), and TransBig ( n = 198) data sets. The three centroid/SSPs used for comparison were described by Sorlie and colleagues in 2003, Hu et al in 2006, and Parker et al in 2009. The results showed a fair to substantial agreement (κ value of 0.238–0.740) between SSPs in each cohort. Only the basal-like class was consistently predicted by each SSP in all the cohorts. However, the prediction of all other classes varied substantially. Excluding the basal-like and luminal A classes, the significance of associations with outcomes of other molecular subtypes varied depending on the SSP used. However, different SSPs produced broadly similar survival curves. This study highlighted the lack of stringent standardization of methodologies, and definitions for molecular subtypes that led to failure of SSPs to reliably assign the same patients to the same molecular subtypes. The study was criticized by the original investigators as well as others in the field. One reason for discordance was probably the use of publicly available data sets with different platforms as pointed out by Perou et al. In practical terms an SSP is valid only when a single platform is used for both training and testing predictions because different protocols for the same gene show measurement bias. One way to deal with bias is to use controls which are often unavailable in public data sets. In these cases, normalization estimates or gene centering must be determined from observations of the test cases, but it defeats the whole purpose of an SSP as one should not go back to all prior samples in order to determine the subtype for the new sample. Despite the criticism of the Weigelt et al study by the original investigators, it is obvious that variations of the methodology (e.g., probe annotation, choice and averaging of probes, and data centering) exert a substantial effect on the assignment of individual samples to the molecular subtypes. Nevertheless, a commercial assay (Prosigna, discussed later) based on intrinsic gene set–based classification is now available for determining breast cancer prognosis.

Molecular Classification of Triple-negative Breast Carcinomas

Triple-negative breast carcinoma constitutes about 15% of breast cancers and is a tumor subgroup that is heterogeneous based on histotype, proteomic, mutational, immune and metabolic profiles. Given the negative HR profile, conventional chemotherapy has traditionally been the major if not sole option for treatment in most cases. The most common histotype seen in this subgroup of tumors is invasive ductal carcinoma. Most of these tumors tend to have high nuclear grades, high mitotic activity, high mutation rates with deficiency in DNA damage repair, and increased immune infiltrate compared to other breast tumor subtypes, and show atypical medullary patterns as described by Livasy et al or no special type (NST) patterns. Morphological overlaps between triple-negative breast carcinomas and tumors associated with germline BRCA1 mutations have also been observed, despite the absence of BRCA mutations in sporadic triple-negative breast carcinomas. While triple-negative breast carcinoma is by definition negative for ER, progesterone receptor (PgR), and HER2, a subset of these tumors express ER-related markers (e.g., androgen receptors). Clinical indolent subtypes of triple-negative breast carcinoma exist, such as adenoid cystic carcinoma, acinic cell carcinoma, low-grade adenosquamous carcinoma, and secretory carcinoma, and some are associated with unique molecular profiles (discussed in further detail in “other molecular subtypes”).

So far no unifying mutational profile has been identified that applies to the entire group of triple-negative breast carcinomas. This tumor group has higher numbers of somatic mutations than luminal breast carcinomas, with main mutations seen involving TP53 (about 50% to 80%) and PIK3CA (about 10% to 20%), with a higher frequency of PIK3CA mutations noted among those that are positive for androgen receptors (AR). Metastatic triple-negative breast carcinomas with PIK3CA/AKT1/PTEN gene alterations were shown to be associated with longer progression-free survival when treated with AKT inhibitors and chemotherapy.

Triple-negative breast carcinomas have been classified in general into basal and nonbasal classes based on their expression of basal markers ( Table 20.1 ). Additional subtypes including basal-like (BL1 and BL2), mesenchymal (M), and luminal androgen receptor (LAR) subgroups have also been identified with different classification systems of gene expression profiling. Other subgroups, such as claudin-low class (further described in “other molecular subtypes [claudin-low tumors]”), mesenchymal stemlike (MSL), and immunomodulatory (IM) subtypes, have also been observed ( Table 20.1 ). The different classification schemes highlight the extent of heterogeneity within triple-negative breast carcinomas, and also help in clinical trial design for assessing treatment response. Different molecular subtypes have been noted to correlate with different response rates to neoadjuvant chemotherapy, with BL1 showing more favorable response and the LAR intrinsic molecular subtype exhibiting a lower response rate.

Lehmann and colleagues performed a meta-analysis of 21 studies comprising 587 triple-negative tumors. The study revealed six stable phenotypes which they referred to as BL1 and BL2, and one IM, one M, one MSL, and one LAR subtype. The BL1 and BL2 subtypes had higher expression of cell cycle and DNA damage response genes, similar to what is observed in basal-like tumors that develop in BRCA1 mutation carriers and are responsive to platinum-based drugs and poly ADP-ribose polymerase (PARP) inhibitors. The BL1 subtype was heavily enriched in cell cycle and cell division pathways and the BL2 subtype displayed unique gene ontologies involving growth factor signaling. The IM subtype showed increased expression for immune signaling genes that substantially overlapped with a gene signature for medullary breast cancer. The M and MSL subtypes were enriched in gene expression for epithelial-mesenchymal transition and growth factor pathways; such tumors would likely respond to PI3K/m-TOR inhibitors and abl/src inhibitors. The LAR subtype included patients with tumors showing AR signaling which is sensitive to bicalutamide (an AR antagonist). Morphoimmunohistological correlation was not performed in this study, but similar groups are identified on morphological and immunohistological analyses of triple-negative tumors. Burstein et al classified triple-negative tumors into four subgroups: basal-like immune activated (BLIA), basal-like immunosuppressed (BLIS), MES, and LAR. It appears that the BLIA class is similar to Lehmann’s IM class, BLIS includes the BL1 and BL2 classes, MES is similar to Lehmann’s M and MSL classes, and LAR is similar in both classifications. Both classifications tend to have prognostic and therapeutic value. Masuda et al reported differential pCR rates after neoadjuvant chemotherapy for each of the Lehmann classes: 52% (11/21) for BL1, 0% (0/8) for BL2, 31% (8/26) for M, 30% (8/27) for IM, 23% (3/13) for MSL, and 10% (2/20) for LAR. This corresponds to excellent response to chemotherapy in typical triple-negative tumors, somewhat attenuated response in metaplastic-type carcinomas, and a general lack of response in low-grade AR+ triple-negative tumors. However, the Masuda et al study highlights the importance of dissecting out which triple-negative breast cancers respond to typical first-line chemotherapy and which tumors require additional targets. BL1 and BL2 are both highly proliferative triple-negative tumors, but BL2 does not respond favorably to anthracycline- and taxane-based chemotherapy.

The role of the tumor immune microenvironment in the biology of triple-negative breast carcinomas has gained much attention in the past 10 years. Analyses of immune-associated gene sets reveal three clusters of triple-negative breast carcinomas, including immunity-low, immunity-medium, and immunity-high, with the immunity-high subgroup harboring a greater extent of tumor-infiltrating lymphocytes (TILs) and carrying an overall better prognosis. Higher levels of TILs have been known to be associated with better response to neoadjuvant chemotherapy immunotherapy, and better prognosis. Adams et al in a recent study examined the expression of PD-L1, CD163, FOXP3, and MCT4 (the latter a marker for glycolytic microenvironment) in triple-negative breast carcinomas, and identified four clusters, with clusters 1 and 2 characterized by high TILs and low PD-L1/FOXP3, and clusters 3 and 4 featuring increased PD-L1, FOXP3, and MCT4 expression. Triple-negative breast carcinomas classified in clusters 1 and 2 were found to show better outcomes than those in clusters 3 and 4.

Combining the existing classification systems of triple-negative breast carcinomas to achieve consensus subtypes is a challenge that needs to be resolved through the integration of different data sources. One such recent study was performed to stratify triple-negative breast carcinomas using multiple layers of data types, including transcriptome, micro-RNA expression, and copy number variants. Further studies to combine other data sources on proteomics, immunogenomics, metabolomics, radiology/imaging, and pathological data would facilitate further subtyping of triple-negative breast carcinomas to guide new treatment and clinical management strategies.

Other Molecular Subtypes (Claudin-low Tumors)

Claudin-low tumors represent a group of triple-negative tumors that have gene expression profiles similar to basal-like breast cancer, but also have some distinct differences. These tumors are characterized by enrichment (much more than basal-like cancers) of epithelial to mesenchymal transition markers, immune response genes, and cancer stem cell markers. These tumors are named as such due to their low gene expression of tight junction protein claudins 3, 4, and 7. Additionally, these tumors demonstrate low to absent expression of the E-Cadherin (ECAD) protein, but are not lobular carcinomas morphologically. Prat et al comprehensively characterized these tumors at the molecular level. Taking tumors from three previously published data sets and 37 new samples ( n = 337) and implementing hierarchical clustering using approximately 1,900 intrinsic genes, the investigators identified the defined intrinsic molecular classes, including claudin-low, which was placed in close proximity to the basal-like subtype. These tumors showed inconsistent expression of basal keratins such as keratins 5, 14, and 17, and low expression of HER2, luminal markers for keratins 18 and 19, ER, PgR, and ER-related genes. The morphological data on claudin-low tumors are scarce but suggest that these tumors are generally high grade and poorly differentiated and often have very intense immune cell infiltrate. Metaplastic breast carcinomas are also thought to be of the claudin-low subtype. However, metaplastic breast cancers can show numerous histomorphological patterns and are therefore expected to be heterogeneous at the molecular level as well. In a gene expression and copy number profiling study of 28 cases by Weigelt et al, only the spindle cell metaplastic carcinomas were of the claudin-low subtype. The metaplastic tumors with squamous and chondroid differentiation were preferentially classified as being of the basal-like subtype. As expected, the prognosis of claudin-low tumors is reported to be significantly worse than that of luminal A tumors (the prognosis of ER– tumors in general is worse than that of ER+ tumors). It is unclear whether these tumors have a worse or better prognosis than basal-like breast cancers. In a microarray gene expression profiling study of 107 triple-negative breast cancers, Jézéquel et al performed unsupervised analysis to identify three different clusters (albeit fuzzy clustering). The first cluster (C1) comprised tumors with luminal-like characteristics with AR expression (i.e., LAR). The second cluster (C2) showed typical basal-like characteristics. The third cluster (C3) included basal-like and claudin-low tumors. The patients belonging to the C3 cluster had better outcomes than those in the C1 ( p = .01) and C2 ( p = .02) clusters. The researchers further examined disease-free survival by pooling this initial cohort with an external cohort of 87 cases (for a total of 194 cases). The C3 (claudin-low) patients again showed better disease-free survival than the C2 (pure basal-like) patients. Additionally, high immune response was a hallmark of the C3 (claudin-low–enriched) cluster. The claudin-low tumors did benefit from routine breast cancer chemotherapy regimens, but it is unclear whether they showed lower or higher pCR rates to neoadjuvant chemotherapy compared to other basal-like breast cancers. Obviously, better and more targeted therapies need to be studied in this specific subgroup of tumors.

Other Molecular Subtypes (Molecular Apocrine Class)

Subsequent to intrinsic gene set–based molecular classification, Farmer and colleagues reported identification of a “molecular apocrine group.” Tumor samples from 49 patients with locally advanced breast cancers were assessed using principal component analysis and hierarchical clustering. Three tumor groups were identified, which were labeled as luminal, basal, and apocrine. The molecular apocrine tumors were described to show strong apocrine features on histological examination, and were positive for AR expression but negative for ER and PgR expression. Further testing revealed that androgen signaling was most active in molecular apocrine tumors and that the tumors were commonly HER2+. It appears that the molecular apocrine group identified by Farmer et al significantly overlaps with the ERBB2 (or HER2-enriched) subgroup of intrinsic gene set–based molecular classification. Extrapolation of these findings to routine diagnostic and IHC analysis suggests that the molecular apocrine tumors are either ER–/PR–/HER2+/AR+ or ER–/PR–/HER2–/AR+. Both tumor groups demonstrate histological evidence of apocrine differentiation in a large percentage of cases. The above assumption is supported by our own study of AR expression in 189 consecutive breast cancers, in which AR was expressed not only in ER+ tumors but also in ER– tumors with apocrine differentiation. The molecular apocrine group is now considered a defined molecular subgroup of triple-negative breast cancers (also known as luminal/apocrine triple-negative breast carcinoma with androgen receptor overexpression). Since these tumors express AR, AR inhibitors such as bicalutamide and enzalutamide are being tested in clinical trials. Histone deacetylase (HDAC) inhibitors are another group of drugs that could specifically work on these tumors. HDAC inhibitors have the potential to turn ER– tumors to ER+ tumors, which can subsequently be treated with antiestrogen therapies.

Molecular Classifications of HER2+ Breast Carcinomas

HER2 positivity, as determined by IHC and/or in situ hybridization (ISH), is seen in about 15% to up to 20% of invasive breast carcinomas. HER2+ breast cancers comprise a heterogeneous group of tumors in terms of their molecular profiles. While most (about 50% to 70%) of HER2+ tumors can be classified as being of the HER2-enriched intrinsic subtype, the remaining HER2+ breast carcinomas have been classified with other intrinsic molecular subtypes, such as luminal A and basal-like subtypes

There could be different reasons for this discrepancy, such as heterogeneity for HER2 overexpression/amplification or some truly HER2– apocrine tumors that are generally negative for ER/PgR but express AR. These triple-negative apocrine tumors likely cluster with the ERBB2 (HER2-enriched) class. Moreover, the triple-negative apocrine tumors appear morphologically more similar to ERBB2 tumors (ER–/PR–/HER2+) than to typical triple-negative basal-like tumors. Farmer et al, who described the molecular apocrine tumors, suggested a simple IHC classification (based on their expression profiling experiments) in which they considered luminal tumors to be AR+/ER+, basal tumors to be AR–/ER–, and molecular apocrine tumors to be AR+/ER–. Other molecular and IHC studies also suggest that molecular apocrine tumors are ER–/PR–/HER2+/AR+ or ER–/PR–/HER2–/AR+ and show significant overlap with the intrinsic molecular class ERBB2. These findings likely explain the inclusion of some clinically HER2– tumors within the ERBB2 intrinsic molecular class. Conversely, Parker et al showed that 11% of ER+ tumors clustered within the ERBB2 molecular class. The exact reason is not known, but as mentioned previously, the clinical cutpoints to define ER positivity (i.e., very low ER-expressing tumors, but considered clinically ER+) are likely the most reasonable explanation for this discrepancy. Although distinct molecular classes are identified by gene expression profiling, there are always some tumors that form a continuum between the two distinct classes.

HER2+ tumors were found to have heterogeneous responses to therapies. Pathological features other than HER2+ status were noted to affect responsiveness to targeted therapy among HER2+ tumors, such as the level of HER2 protein expression and the level of HER2 gene amplification, with higher HER2 protein expression and amplification levels being associated with higher complete pathological response rates. Immune signatures and PIK3CA mutational status within the tumors also play a role, with patients having low immunogenicity in tumors responding better to HER2-targeted therapies, and patients with PIK3CA exon 9 mutation showing decreased sensitivity to HER2-targeted monoclonal antibody treatment. As such, HER2-enriched molecular status has been studied by some to be a possibly better predictor to HER2-targeted therapy than HER2 IHC or ISH positivity. In our routine practice, we have commonly observed and previously reported that coexpression of HR significantly affects response to neoadjuvant chemotherapy plus anti-HER2 therapy. The expected pCR to neoadjuvant chemotherapy plus anti-HER2 therapy in ER+/HER2+ therapy ranges from 30% to 40% (higher with the addition of pertuzumab) compared to 55% to 70% in ER–/HER2+ tumors (again, higher with the addition of pertuzumab). Additional investigation of the role(s) of immune and PI3K pathways in tumors’ responses to HER2-targeted therapy is warranted to optimize combined therapeutic regimens that target the PI3K and immune checkpoint pathways.