Translational Genomics: Practical Applications of the Genomic Revolution in Breast Cancer

Technical and analytic developments underpin major advances in our understanding of breast cancer biology over the last two decades. The application of gene expression arrays in the early 2000s led to a refined molecular classification, the development of prognostic assays, and a better characterization of the contribution of the tumor microenvironment (1, 2). First steps toward translating molecularly defined breast cancer subsets came from efforts using gene expression–based signatures, which have led to multiple commercially available and clinically useful assays in hormone receptor–positive breast cancer, and are continuing to be developed for multiple clinical/translational applications today. Discussing RNA-based assays is beyond the scope of the current review but was the topic of a recent CCR 20^th Anniversary Commentary (2). The genomic revolution later saw the emergence of commercially available massively parallel sequencing approaches that permit the comprehensive profiling of an entire cancer's genome, all coding genes (exome), or a selection of genes (targeted sequencing). DNA sequencing technologies offer the potential to deliver personalized medicine by matching appropriate targeted therapeutics with unique molecular aberrations within an individual's cancer. Furthermore, massively parallel sequencing can be used to identify biomarkers associated with response to experimental targeted therapies, both in clinical trials and in vitro.

Here we consider five practical applications of the genomic revolution, summarizing in each case what we have learned thus far and considering the challenges that remain for future endeavors (Fig. 1). Key messages include the following: (i) the driver landscape of breast cancer includes many rare cancer genes and an unexpected level of inter- and intratumoral driver mutation heterogeneity; (ii) understanding the mutational processes operating throughout a cancer's lifetime may identify therapeutic vulnerabilities; (iii) many genomic features align with traditional molecular and histologic subtypes that can aid classification; (iv) cancers demonstrate subclonal diversification that may underlie primary and acquired treatment resistance; and (v) a range of patterns of metastatic dissemination are observed reflecting clonal evolution.

To date, almost 100 high-probability breast cancer driver genes have been identified through the application of computational approaches to massively parallel sequence data generated from more than 4,000 primary breast cancer samples (Supplementary Table S1; refs. 3–18). Cancer genes have been identified based upon somatic mutation recurrence rates that exceed expectation after taking into account features such as background mutation rate and spectrum, nucleotide context, enrichment in phosphorylation-specific regions or tertiary protein structures, and the ratio of nonsynonymous to synonymous mutations (19–24). Various alternative methodologies exist such as those that adopt subnetwork analyses to identify groups of potential cancer genes by incorporating pathway or gene interaction information, and these may be particularly useful for identifying very rare driver genes (25–28).

Saturation analyses predict that based on the number of samples now sequenced, most cancer genes implicated in 2% or more breast cancers will have been identified (29). However, we expect that many more low-frequency cancer genes remain to be discovered, because there is emerging evidence, as discussed below, that metastatic, pretreated, male, and special histopathologic subtypes of cancers are genomically distinct from the “general” breast cancer population (Tables 1 and 2; Supplementary Table S1; refs. 6, 10–18, 30–37). Rare “populations” are often the most clinically challenging to treat, and further focused driver discovery exercises in these groups are warranted.

Driver alterations refer to events that are predicted to inactivate tumor suppressor genes or activate oncogenes—common mechanisms in the former include homozygous deletions, truncating point mutations, or disrupting rearrangements and in the latter, focal amplifications and activating missense mutations, typically within mutational hotspots. A recent genome-wide analysis of 560 breast cancers (9) reported that the average cancer contains three such driver alterations (range = 0–14) and, consistent with other series, confirmed that only a handful of cancer genes are mutated in more than 10% of breast cancers (3–10). The most frequently implicated cancer genes (and their frequencies) across the 560 breast cancers are TP53 (41%), PIK3CA (30%), MYC (20%), PTEN (16%), CCND1 (16%), ERBB2 (13%), FGFR1 (11%), and GATA3 (10%; ref. 9). Because most cancer genes are mutated in less than 5% of cancers, considerable interpatient heterogeneity exists (4, 7). By example, whole-exome sequencing of 100 breast cancers identified driver point mutations and copy number changes in 40 genes in 73 combinations, indicating that at the driver gene level, most primary breast cancers are distinct (4). The wide range of possible driver combinations within an individual cancer will demand systematic approaches to facilitate target prioritization and guide combination therapy approaches if we reach a position where a large number of effective therapeutic targets are available.

The majority of driver events detected in a primary tumor are clonal, being present in all tumor cells. However, subclones are common, and these can harbor driver events that alter some of the most common and most frequently clonal breast cancer genes including TP53, MYC, PIK3CA, and HER2, in addition to rarer cancer genes such as FGFR2 and RUNX1 (3, 35, 36, 38, 39). In a multiregion sampling study, we identified subclonal, potentially targetable alterations in 13 of 50 primary breast cancers and demonstrated that subclones can be spatially separated, posing a challenge to representative sampling (36). In a quarter of these cases, we identified parallel evolution, with the same cancer gene mutated multiple times in different subclones in the same tumor, indicating that particular subclonal aberrations within an individual cancer may be under strong selection (36). Highlighting the clinical relevance of such a finding, in one case of metastatic breast cancer, parallel genetic loss of PTEN in multiple metastatic sites was documented after treatment with the PI(3)Kα inhibitor BYL719 (40). In this case, the emergence of a convergent PTEN-null treatment–resistant phenotype provided strong rationale for targeting the subclonal alterations; however, in general, it is not yet clear when and if we should target subclonal alterations or if a greater benefit will be derived from targeting clonal alterations that are present in every cancer cell.

Determining the range of driver events operating within an individual breast cancer depends, to some extent, on the sequencing technology employed (Fig. 2). Targeted approaches can provide (at lower cost than whole-genome sequencing) high-depth sequence coverage, high sensitivity, and accurate quantification of point mutations, overcoming, to some extent, technical challenges such as low tumor cellularity and driver mutation subclonality (ref. 41; Fig. 2). However, targeted approaches make prior assumptions about the range of relevant cancer genes. Furthermore, breast cancers are characterized by frequent and often complex structural variation, and the precise nature of structural variation can only be captured using genome-wide sequencing (Fig. 2; ref. 42). For example, a recent study revealed that rearrangement breakpoints disrupt tumor suppressor gene footprints more frequently than expected, contributing to the recessive cancer gene landscape, with the most frequently affected genes being PTEN, RB1, and ARID1B in 7%, 5%, and 4% of the cases, respectively (9). The same study revealed that recurrent fusion genes and noncoding driver mutations are uncommon in primary breast cancers (9). The full contribution of structural variation to the driver landscape of breast cancers remains to be determined. Selecting the most appropriate technology for incorporation in clinico-genomic trials is of paramount importance.

Defining the targetable landscape of breast cancer remains in its infancy. High-throughput functional validation approaches are indicated to determine the significance of potential cancer genes and driver mutations (43, 44). Importantly, the clinical viability of even some of the most common cancer genes (present in more than 10% of cancers) as therapeutic targets remains as yet unproven, with only ERBB2 and PIK3CA (45) associated with objective responses to targeted therapies and only the former demonstrating validated and independent prognostic and predictive value (46). TP53, GATA3, and MYC have diverse transcriptional activities, and attempts to target them both directly and indirectly have proved difficult for various reasons, including lack of specificity and efficiency (47–51). These targets remain an intense area of research. For example, clinical trials assessing the activity of PIM kinase inhibitors in MYC-amplified triple-negative breast cancers are now keenly awaited following promising preclinical work that identified a targetable synthetic lethal interaction between PIM1 and MYC (52).

It is important to consider that the presence of an apparent target does not necessarily predict for response. Recent phase III studies confirmed a survival advantage associated with the use of mTOR inhibitors (53) and CDK4/6 inhibitors (54) in advanced estrogen receptor (ER)–positive breast cancers; however, to date, biomarkers to predict response to these agents have not been identified. In the PALOMA 1 trial, CCND1 amplifications or loss of p16 failed to predict for a response to the CDK4/6 inhibitor palbociclib (55). In a retrospective analysis of the BOLERO-2 trial, progression-free survival benefit was observed in association with the mTOR inhibitor everolimus irrespective of the presence or absence of genomic alterations within PIK3CA, CCND1, and FGFR1 (56). However, the extent of benefit from everolimus varied according to the exon specific mutation (exon 20 versus exon 9), indicating that the definition of a target may be more nuanced than initially appreciated.

Overcoming the significant challenges of defining and ultimately targeting the driver landscape of breast cancer will demand large-scale, international, novel clinico-genomic screening and clinical trial strategies, as has been discussed comprehensively elsewhere (57). It is also clear that, since many driver alterations are ultarare in breast cancer, studies from single cases or small case series will also be extremely valuable, particularly for exploring extreme response or therapeutic resistance (58).

Whole-genome sequencing identifies thousands of mutations within each cancer, and although the majority of them are passenger events, the patterns that they form, termed “mutational signatures,” can reveal the mutational processes that shaped that cancer's evolution (59). A recent analysis of 560 breast cancer whole genomes identified 20 distinct mutational signatures (9). Typically, multiple mutational processes are identified within each individual breast cancer, and their relative contributions may vary over time (36, 60).

A comprehensive overview of mutational processes in breast cancer is presented in another article in this CCR Focus section (61). In the context of “practical applications,” the fundamental principle is that understanding the processes that generate mutational diversity, the substrate of natural selection, could ultimately result in strategies to oppose cancer development and progression (62). However, for most mutational signatures, both etiology and direct clinical utility are at this time unclear. At present, the mutation signatures with the greatest potential clinical benefit are those that identify deficiencies in the homologous recombination repair pathway and may predict for sensitivity to platinum salt chemotherapies and PARP inhibitors (63–68). Such signatures could extend the number of patients for whom such agents may confer benefit, beyond those with germline or somatic mutations or promoter methylation within BRCA1/2 and other BRCA-related genes (9, 69, 70), and has been used to confirm that an ER-positive tumor arising in a BRCA1 carrier is related to the germline lesion rather than being a sporadic cancer (71).

Despite considerable intertumoral heterogeneity, certain genomic patterns align with the molecular and histologic subtypes. For example, luminal-like/ER-positive breast cancers typically harbor mutations in PIK3CA, GATA3, or MAP3K1 (6–8). The most commonly mutated cancer gene in basal-like/triple-negative breast cancers is TP53, and this usually contains truncating (frameshift or nonsense) mutations (72). In contrast, when TP53 mutations occur in luminal-like cancers, they are usually missense mutations (8, 10, 73). Characteristic genomic changes also underlie histologic subtypes. Recently, several consortia have genomically characterized more than 1,225 invasive lobular breast cancers (6, 10, 13, 17, 18). Lobular tumors typically express ER and lack HER2 amplification, and represent the second most frequent histologic subtype, accounting for up to 15% of all invasive breast carcinomas (74). Lobular tumors display a distinct mutational pattern compared with their ER-positive/HER2-negative ductal counterparts, including a higher frequency of CDH1 and PTEN loss and AKT1, ERBB2, ERBB3, and TBX3 mutations (6, 10, 13, 17). Furthermore, ER-positive tumors of lobular and ductal type have a strong tendency toward FOXA1 or GATA3 mutations, respectively, indicating that distinct genomic mechanisms of ER activation may exist in these distinct pathologies (6, 17, 75). Distinct genomic patterns are also characteristic of rare tumor types, including adenoid cystic, apocrine, medullary, mucinous, metaplastic, micropapillary, and neuroendocrine tumors, and are detailed in Table 1 and Supplementary Table S1 (10–12, 15, 16, 18, 32, 33, 76).

We continue to extrapolate from global breast cancer populations to make clinical treatment decisions for special histologic subtypes. Now, there is growing evidence of molecular differences between these subtypes. We should, therefore, acknowledge both molecular and histologic subtypes as relevant pathologic covariates in future clinical and translational research. In addition, because certain potential therapeutic targets are enriched in some subtypes, such as ERBB2 and ERBB3 mutations in lobular carcinomas, one could imagine that conventional pathology could be used to enhance mutational screening for clinical trial enrollment.

Clinically overt metastasis to distant organs is currently incurable. The most informative sequencing approaches exploring this fatal transition compare primary tumor and metastatic deposits from the same individual (Supplementary Table S1). Studies focused on coding mutations reveal that most of the driver events within the primary tumor are shared with associated metastatic deposits (77–80). This indicates that the primary tumor may be a good surrogate for the relapse-seeding clone, which is the target for adjuvant therapies that have the potential to improve cancer cure rates (79, 80). For example, in a study of single metastatic deposits from 33 individuals, 87% of driver mutations and 86% of copy number aberrations were also detected in the primary tumor (79). A recent single-cell sequencing study of disseminated tumor cells supported this linear model of late dissemination from the primary tumor (81). Metastatic deposits generally harbor additional private mutations, often including known driver mutations, reflecting ongoing subclonal diversification after dissemination (36, 77–80, 82).

A handful of published studies have started to unravel the genomic basis of breast cancer progression and define the temporal ordering of mutations through the analysis of the primary tumor and multiple metastatic deposits acquired during “rapid” autopsy (Supplementary Table S1; refs. 40, 77–80, 82–84). In this context, a recent survey conducted in Australia to evaluate patients' attitudes toward rapid autopsy and tissue donation revealed that 87% were willing to donate tissue at death (85). So far, multiple tumor deposits from a total of 17 breast cancer patients who underwent autopsy have been investigated at the mutational and copy number level (40, 80, 83, 84). Although each patient reported in these studies has her own history of treatment and disease evolution, several messages emerge. First, given the observed primary/metastasis and intermetastasis genomic heterogeneity, there is no single tumor sample that recapitulates the complete disease, that is, all the genomic alterations present in all samples from the patient (40, 80, 82–84). Second, both linear and parallel evolution from the primary tumor are observed (Fig. 1; refs. 80, 83, 84). In some patients, there is only one successful seeding event that gives rise to a so-called metastatic precursor that then seeds the other metastases, possibly in a cascading manner (Fig. 1). In other patients, multiple successful seeding events arise from the primary tumor. Finally, evidence is accumulating that metastatic cross-seeding occurs between established metastases (83, 84).

A potential limitation of the studies that investigate metastasis using a targeted sequencing approach is that the cancer genes present in these panels were derived almost exclusively from the study of treatment-naïve primary tumors and, consequently, may fail to identify genes enriched or specific to metastatic, relapsed, or pretreated cancers. Some preliminary evidence from metastatic tumors suggests indeed that different drivers may be associated with metastasis, and that these are drawn from an even wider range of cancer genes than in primary tumors (Supplementary Table S1; refs. 30, 34, 79, 86). It is conceivable that some genes only act as drivers in nuanced situations, for example, in resisting a specific anticancer therapy, and in specific cancer subtypes, so studies should be powered to identify rare resistance mutations, a theme that will be explored in further detail later in this review (34, 87–89).

Extensive work remains within this field to determine when and what driver mutations are most relevant to the application of personalized medicine. Exhaustive sampling of all metastatic deposits is not practical, so if metastasis-specific mutations are clinically important, representative sampling strategies need to be developed. Options include sampling areas of progressive disease after treatment interventions, particularly when a mixed response is seen, or alternatively, using “liquid biopsies” (refs. 90–92; Supplementary Table S1). In one study, sequencing of circulating tumor DNA (ctDNA) successfully identified changes in the levels of subclonal mutations as the relevant metastatic sites responded or progressed, providing proof of principle for the role of a liquid biopsy to monitor treatment response (90). Such an approach can also use ctDNA to predict minimal residual disease and relapse, potentially extending the window of opportunity for applying adjuvant therapies to further improve cancer cure rates (93).

Changes in subclonal structure have been demonstrated across treatment interventions using cytotoxic and endocrine therapy and targeted therapies (35–37, 40, 94). Several genomic mechanisms of resistance to various therapies have recently been uncovered, with mutations in ESR1, the gene encoding the ER, being the most described and studied in the context of endocrine resistance (34, 87, 95–98). Mutations cluster in the ligand-binding domain of ESR1 and activate ER in the absence of the estrogen ligand. These mutations have been detected almost exclusively in the context of prior endocrine therapy, typically with aromatase inhibitors, and more frequently where sensitivity to prior endocrine therapy was observed (98), suggesting that these mutations are associated with acquired, but not primary, endocrine resistance. A recent large-scale initiative, the Metastatic Breast Cancer Project, characterized over 140 metastatic biopsies from endocrine therapy–treated patients and identified several acquired mutations, in addition to ESR1, that may be involved in endocrine resistance, including AKT1, ERBB2, KRAS, and RB1 alterations (99). There is also preliminary evidence from cell-line and patient-derived xenograft models that CDK4/6-inhibitor resistance may be mediated by loss of RB1 or amplification of CCNE1 or CDK6 (100, 101).

Treatment resistance might be complicated by heterogeneity, for example, multiple metastatic deposits from an individual may exhibit parallel evolution of genomic resistance, whereas other metastases may not have any genomic mark of resistance, potentially explaining the heterogeneous response to treatment observed in the clinic (40, 83, 96–98, 102).

Interrogating ctDNA represents an attractive option for identifying multiple mechanisms of resistance in different metastases. This is the strategy adopted by the recent UK plasmaMATCH trial (ISRCTN16945804), a primary aim of which is to determine whether plasma-derived ctDNA is an acceptable alternative to tumor biopsies for mutation identification in metastatic breast cancer. A secondary aim is to determine whether there is a clinical benefit from therapeutic targeting of mutations identified in ctDNA. However, open questions remain regarding the contribution of tumor deposits to ctDNA. This contribution could be dependent on the metastatic organ, cancer proliferation rate, histology, size of the metastasis, and response to therapy. For instance, it has been demonstrated that brain metastases do not or poorly contribute to ctDNA in plasma and that ctDNA from cerebrospinal fluid would offer a better representation of the genomic alterations of brain metastases (103).

It is also important to perform comprehensive assessment of the various pathways activated within an individual cancer rather than seeking a single mutation in isolation (104). For example, only a fraction of patients receiving selective PI3Kα inhibitors respond, and nonresponders tend to demonstrate an increase in ER-dependent transcription that can be overcome in some cases by combining the PI3Kα inhibitor with an anti-estrogen such as tamoxifen or fulvestrant (105).

Although the genomic revolution has profoundly broadened our understanding of the number and nature of somatic alterations within breast cancer genomes, the clinical benefit of these discoveries remains modest. The feasibility of DNA sequencing to direct the management of metastatic breast cancer through a personalized cancer medicine approach has been examined in two prospective molecular screening studies: SAFIR01/UNICANCER and MOSCATO, in the years 2011 to 2012 (106, 107). The two studies included 423 and 129 patients, and identified potentially targetable alterations in 40% and 46% of patients with the ability to personalize therapy in 13% and 23% of cases, respectively. Of the patients who received targeted therapy, 9% and 20% had a partial response and 21% and 56% had stable disease, respectively. Therefore, 3% and 15% of patients originally screened in these two studies derived clinical benefit from a targeted therapy, but in the absence of a control arm, we cannot infer if the benefit exceeded that of standard of care. It is worth noting that the SHIVA trial, which randomized 293 patients with potentially targetable alterations to molecularly targeted versus standard therapy and found no improvement in outcome but higher toxicity, involved last-generation targeted agents (108). The effect of a personalized cancer medicine approach on survival or, indeed, any other outcome remains to be determined within randomized trials using validated biomarkers and modern targeted drugs.

Although this review has focused on genomic sequence data, it is clear that tackling breast cancer, a complex set of diseases, will require integrated assessment of a wide range of data types, including pathologic characterization (such as histologic subtyping and assessment of the cellular composition), transcriptomics, epigenetics, and proteomics, in addition to genomics (6, 10, 109).

Most large driver discovery studies published to date include primary, treatment-naïve cancers (Supplementary Table S1), and around 80% of these are likely to have been cured using existing treatment strategies. Moving forward, focus should be placed on understanding the mechanisms of both primary and acquired treatment resistance and the genomics of metastasis that is almost universally fatal. This should include using genomics to identify primary breast cancers most at risk of relapse and to characterize disseminated tumor cells that are capable of metastasis formation.

We believe that these goals can be achieved through the joint results from many large-scale, prospective initiatives that aim to better characterize metastases and link molecular findings with meticulously annotated clinical and treatment information. These initiatives include the AURORA program, the Metastatic Breast Cancer Project (https://www.mbcproject.org/users/new), the UK plasmaMATCH (ISRCTN16945804), as well as the various rapid autopsy programs (99, 110, 111). However, given the extensive intertumoral heterogeneity of breast cancer, we must not dismiss the power of “n of 1” studies and small case series that could deliver key insights into ultrarare alterations and associated treatment resistance and response.

A large amount of genomic data from several thousand breast cancers has already been generated, and further value can almost certainly be extracted from this. For example, a combined analysis, although challenging on account of the heterogeneity of the datasets, would have much higher power to detect rare mutations than any individual study to date. Supporting such a collaborative approach is the AACR Project Genomics Evidence Neoplasia Information Exchange (GENIE; http://www.aacr.org/Research/Research/Pages/aacr-project-genie.aspx#.WH6TMnCFNOC), an international data-sharing project that aims to collate genomic and clinical data from tens of thousands of patients in multiple institutions.

Great enthusiasm accompanied the initial years of the genomic revolution as a flood of novel insights into the breast cancer genome were presented in quick succession. We now have a clearer idea not only of the potential clinical applications of genomics but also of the scale and nature of the challenges that must be surmounted to bring maximum benefit to breast cancer patients. It is clear that delivering personalized breast cancer treatments to all patients in the future will only be possible if we focus on collaborative and integrative approaches now.

No potential conflicts of interest were disclosed.

This work was supported by a grant from Les Amis de Bordet (to C. Desmedt).