Expressed Gene Fusions as Frequent Drivers of Poor Outcomes in Hormone Receptor–Positive Breast Cancer

The majority of primary breast cancers express estrogen receptor (ER) and progesterone receptor (PR) and are responsive to therapies that target the estrogen response pathway. Thus, ER and PR are favorable prognostic determinants and actionable therapeutic targets (1, 2). Nevertheless, a substantial minority of women treated for primary hormone receptor–positive (HR⁺) disease exhibit hormonal therapy resistance, evidenced by the eventual development of metastatic disease. The pathways that confer hormonal therapy resistance have been studied extensively, yet our understanding of them remains relatively limited. Most notable among these are growth factor signaling networks that converge on the PI3K and MAPK pathways. Somatic mutations affecting these pathways are hallmarks of HR⁺ breast cancer but are found in only a subset of tumors, suggesting that additional, undetected genetic events contribute to this disease (3).

The most frequent genetic alterations in primary HR⁺ breast cancer are mutation or amplification of the PI3K catalytic subunit gene PIK3CA (30%–40% of tumors), mutation of the TP53 tumor suppressor (≤30% of tumors), and amplification of the EGF family receptor gene ERBB2 (HER2; approximately 25% of tumors; ref. 3). These and other, lower prevalence mutations occur in overlapping subsets of tumors, and consequently many tumors lack recognizable drivers. The recent identification of mutations in ESR1 (encoding ERα) almost exclusively in treatment-refractory metastatic disease has suggested that metastases may harbor an additional proportion of disease-relevant mutations (4, 5). But even taken together, the genetic findings from standard exome sequencing and copy-number analyses are likely to represent at best a limited and incomplete picture of relevant genetic events in these tumors.

A potential group of relevant genetic events likely to be overlooked by exome-restricted DNA sequencing and traditional gene copy-number assessment are genomic rearrangements. Although whole-genome sequencing (WGS) can in theory detect many of these events, the complexity and “signal-to-noise” ratio in such analyses poses challenges, and many such rearrangements are not expressed as RNA (6). As expected, transcriptome-based analyses have also revealed the presence of intergenic fusions in breast cancer specimens (7–11). Routine detection of these events remains a challenge, however, and consequently their overall prevalence and potential clinical significance remain essentially unknown.

We recently reported the development and validation of a methodology, anchored multiplex PCR (AMP), for rapid and efficient targeted gene rearrangement detection (12). Here, we apply AMP systematically to cohorts of patients with early-stage and advanced HR⁺ breast cancers. Through complementary FISH analysis and functional studies, we reveal the frequent presence of novel, biologically and clinically relevant fusions involving potentially actionable genes and pathways in advanced HR⁺ breast cancer. The poor outcomes associated with these genetic lesions underscore the significance of our findings and the likely implications for breast cancer prognostication and treatment.

AMP uses a mixture of tumor cDNA and genomic DNA as starting material. The assay has the major advantages that it does not require a priori knowledge of gene fusion partners and is compatible with low nucleic acid input from formalin-fixed, paraffin-embedded (FFPE) specimens (12). By using a nested, one-sided PCR approach AMP also allows identification of individual (pre-PCR) molecules, thereby mitigating PCR artifacts and facilitating quantitative gene expression analysis (Fig. 1A; ref. 12). We designed the AMP assay to detect fusions involving 54 solid tumor–relevant genes (Supplementary Table S1), and in our proof-of-concept study, we demonstrated a sensitivity of 100% and specificity of 100% compared with reference assays for gene rearrangement detection from FFPE samples (12).

We then applied AMP to tumor specimens from 110 sequential patients with advanced HR⁺ breast cancer who had presented to our clinic and had undergone clinical somatic tumor genotyping (Clinical Genotyping Cohort). This otherwise unselected patient cohort is representative of patients presenting with advanced HR⁺ disease, as most (>80%) had initially presented as stage II and higher, and the majority who presented with primary disease received adjuvant hormonal therapy (>90%) and/or chemotherapy (57%). Median age (49.5) was younger than average for such patients but is consistent with the demographic seen at an academic center (Supplementary Table S2). Remarkably, 14 of these patients harbored high-confidence intergenic fusions (Fig. 1B; Supplementary Table S3). Notably, the majority of these involved exon–exon fusions, including one case with an in-frame fusion of ESR1 to CCDC170, which has been previously reported to occur in <1% of HR⁺ breast cancers, and another with a fusion involving BRAF and the SND1 gene, previously described in lung cancer (Supplementary Table S3; refs. 13, 14). The major identified fusion junction sequences in all cases involved at least one precise exon boundary, thereby defining the detected sequences as processed cDNA (Fig. 1C and D).

Because we wished to determine how often these fusions were present in primary tumors and the initial cohort represented a mix of primary and metastatic specimens, we next analyzed a second cohort of patients with advanced HR⁺ breast cancer (N = 63) for which both primary and matched metastatic specimens were available (Matched Primary/Metastasis Cohort; see Methods for cohort enrollment details). The clinical features of this patient cohort are very similar to those of the Clinical Genotyping Cohort, other than a slightly higher median age (58.5) and increased use of tamoxifen versus aromatase inhibitors in the adjuvant setting (Supplementary Table S2). The latter was expected, as a subset of these patients were enrolled at international sites where, until very recently, tamoxifen remained the most prescribed hormonal therapy agent. In this cohort, 10 of 63 patients were found to harbor a fusion in either the primary or metastatic specimen, 5 of which were detectable in the primary tumor (Supplementary Fig. S1A). Collectively, fusions were detected in 24 of 173 patients (14%), including 11 primary tumors (summarized in Fig. 1B). Notably, fully one-third of these fusions (8/24) involved ESR1, and multiple additional recurrent fusion targets were found in these cohorts, including two fusions each involving AKT3, NOTCH1, PRKCA (encoding protein kinase C), and BRAF (Supplementary Table S3). Among the most notable alterations identified were in-frame fusions involving the key oncogenic drivers PIK3CA, AKT3, RAF1 (also known as CRAF), and ESR1, in each case fused to a distinct partner gene (Fig. 1C and D; Supplementary Fig. S1B and C).

In order to characterize these samples in more detail, we also performed mutational analysis using a site-specific, clinical laboratory assay that detects 120 recurrent mutations in 16 key cancer genes (Supplementary Table S4; ref. 15). We identified somatic driver mutations in 34% (59/173) of these patients, the majority involving PIK3CA as anticipated (Fig. 1B). Discordance between mutations detected in matched primary and metastatic specimens was noted as previously reported (16), and such discordance was also observed for fusions (Supplementary Fig. S1A). Interestingly, we did not observe a statistical association between the presence or absence of mutations and that of fusions in these cohorts (Fig. 1B).

As a next step to credential the potential relevance of these fusions, we sought to confirm the presence of the corresponding chromosomal rearrangements using an independent methodology, FISH. Approximately half of the detected intergenic fusions we identified involved genes on different chromosomes (Supplementary Fig. S2A). And among those involving genes on the same chromosome, in most cases the direction of transcription and/or centromere–telomere orientation predicted a genomic rearrangement detectable by FISH. We used a “break-apart” FISH strategy, incorporating different color probes 5′ and 3′ of the interrogated genetic loci, so that genomic rearrangement is evidenced by a separation of the normally juxtaposed red and green signals (Fig. 2A; ref. 17). The FISH approach also allowed us to ask whether these rearrangements were clonal (appearing in most or all identifiable tumor cells within a specimen), and whether they were present in primary tumors and metastases from the same patient.

FISH analysis confirmed a rearrangement of the respective locus in each AMP-positive tissue tested. Thus, FISH “break-apart” of the RAF1 gene was evident in two distinct metastatic specimens from a patient harboring a CTNNBL1–RAF1 fusion detected by AMP (Fig. 2B). Break-apart of AKT3 was evident in both the primary and two metastatic specimens from a patient harboring an RPS6KC1–AKT3 fusion detected by AMP (Fig. 2C). Notably, in this case, FISH identified both rearrangement and high-level amplification of the AKT3 locus in the primary and metastatic lesions (Fig. 2C). Break-apart of ESR1 was evident in a primary tumor and multiple metastases from a patient whose tumors all demonstrated an ESR1–COA5 fusion upon AMP analysis (Fig. 2D). Additionally, the previously described ESR1–CCDC170 fusion detected by AMP was confirmed upon FISH analysis for ESR1 (Supplementary Fig. S2B). Finally, a fusion involving BRAF and the SND1 gene was detected by AMP and was validated by FISH in two independent patients (Fig. 2E; ref. 14). We then developed IHC assays for RAF1 and AKT3 and observed high-level protein expression of the respective driver genes both in primary and metastatic tumors expressing these fusions (Supplementary Fig. S2C and S2D). Notably, fusion-negative breast cancer cases show little or no expression of these proteins. Collectively, these findings support fusions detected by AMP as bona fide, expressed genomic aberrations.

Several novel kinase-associated fusions we identified juxtapose a fully intact catalytic domain of the respective kinase to the partner gene (Supplementary Fig. S1B). In order to assess the potential contribution of these fusion proteins, we first analyzed their effects in nontransformed mammary epithelial cells (MCF10A). We synthesized cDNAs encoding predicted fusions involving PIK3CA, RAF1, and AKT3 and introduced them via lentiviral vectors, demonstrating stable expression of proteins of the predicted molecular weight in each case (Fig. 3A). Furthermore, expression of these altered kinases was sufficient to activate mTORC1 signaling in these cells. This was evidenced by increases in phosphorylation of downstream substrates, including p70 S6 kinase (S6K1) and its substrate ribosomal protein S6 (RPS6), which was most evident in the absence of serum (Fig. 3B).

To assess the phenotypic effects of pathway activation by these kinase fusions, we determined their effect on acinar growth in three-dimensional (3-D) basement membrane (Matrigel) culture. Parental MCF10A cells produce stereotypical hollow spheres (acini) comprised of polarized cells in this setting, and deregulation of acinar morphology following introduction of various genes is known to correlate strongly with oncogenic properties of the respective genes in human breast cancer (18, 19). We found that expression of each fusion induced significant abnormalities in 3-D culture, and, furthermore, each produced a distinct disruption of the typical acinar growth. The PIK3CA fusion showed the strongest effect on cell growth, acinar morphology, and polarization, evidenced by the appearance of large, disordered structures (Fig. 3C). In contrast, all three fusions were associated with a significant increase in filled acinar structures (Fig. 3C). The filled acinar phenotype indicates defective apoptosis, because a well-described apoptotic program is known to be required for acinar luminal clearance in this setting (20).

We next used an independent assay to detect the apoptosis program in this context by staining acini with ethidium bromide, which is nonpermeable to viable cells. We found, as predicted, that control 3-D acinar structures exhibited ethidium-stained nuclei, indicating cell death, commensurate with luminal hollowing (Fig. 3D). In contrast, structures expressing the AKT3 fusion demonstrated near-complete absence of ethidium-stained cells, which correlated with the highest proportion of filled acinar lumens (Fig. 3D). The RAF1 and PIK3CA fusions conferred a nonsignificant trend toward decreased ethidium staining in 3-D culture (Fig. 3D; Supplementary Fig. S3), consistent with their effects on acinar filling (Fig. 3C). Thus, these kinase fusions activate phosphoprotein signaling and induce oncogenic deregulation of cell growth, polarity, and survival in mammary epithelia.

We then sought to uncover a specific mechanism of fusion-mediated kinase deregulation. We focused on the fusion involving RPS6KC1 and AKT3, as the encoded protein had the strongest effect on cell survival in 3-D culture. RPS6KC1 encodes a pseudokinase whose amino terminus contains a PX domain, an established phosphoinositide (PI) binding motif (Fig. 4A; ref. 21). AKT3 encodes an established oncogene that is frequently subject to genomic amplification in gliomas and other tumors (22). The identified fusion juxtaposes the PX domain of RPS6KC1 to a nearly intact AKT3 protein. Importantly, PX domains typically bind a broader variety of PI species than does the pleckstrin homology (PH) domain of AKT3 (21). Thus, we hypothesized that the kinase fusion might be rendered constitutively active by virtue of membrane recruitment in the absence of PI3K activation and subsequent PIP3 generation (Fig. 4A). To test this hypothesis, we performed a direct biochemical and functional comparison of the RPS6KC1–AKT3 fusion to wild-type AKT3. Expression of RPS6KC1–AKT3 and AKT3 at comparable levels in MCF10A indeed revealed increased basal phosphorylation of the fusion in the absence of serum (Fig. 4B; Supplementary Fig. S4A). Correspondingly, we observed elevated phosphorylation of the direct AKT substrates FOXO3A and Tuberin (TSC2) and of the downstream substrate RPS6 following expression of the fusion compared with wild-type AKT3 (Fig. 4B; Supplementary Fig. S4B).

In order to test directly the prediction that RPS6KC1–AKT3 was constitutively membrane associated, we fractionated serum-starved or stimulated cells expressing the fusion or wild-type AKT3. We observed the expected weak membrane localization of wild-type AKT3 in starved cells, followed by marked membrane recruitment with serum stimulation. In contrast, the fusion protein was constitutively membrane localized in multiple cell types and demonstrated no further membrane recruitment upon serum stimulation (Fig. 4C; Supplementary Fig. S4C and S3D). Consistent with these findings, expression of RPS6KC1–AKT3 induced nuclear exclusion of FOXO3A, which is known to be induced by phosphorylation, in MCF10A (Fig. 4D). The effects of this novel kinase were further corroborated in 3-D membrane culture, as the expression of the fusion produced large, disordered acini with filled lumens, which appeared identical to those described previously following expression of a constitutively active (myristoylated) AKT (Fig. 3C; ref. 23). Thus, RPS6KC1–AKT3 fusion produces a constitutively membrane-associated and activated kinase with oncogenic properties.

We hypothesized that these novel fusions, as potential drivers of hormone-dependent breast cancers, would alter hormone dependence and the response to hormonal therapy in cell-based HR⁺ breast cancer models (24, 25). We therefore expressed three novel kinase fusions involving RAF1, AKT3, and PIK3CA in estrogen-dependent MCF7 and T47D breast cancer cells (Fig. 5A; Supplementary Fig. S5A). Both RAF1 and AKT3 fusions substantially activated downstream phosphoprotein signaling in these cells (Fig. 5B). Further demonstrating the activity of these kinases in this context, expression of the AKT3 fusion conferred resistance to a small-molecule AKT inhibitor (MK-2206), whereas expression of the RAF1 fusion conferred resistance selectively to the clinical MEK1/2 inhibitor trametinib (Supplementary Fig. S5B; refs. 26, 27).

Altered signaling induced by the different fusions produced remarkably distinct effects on hormone-dependent proliferation. Although MCF7 and T47D cells normally undergo complete arrest in the absence of estrogen, expression of the AKT3 fusion consistently conferred estrogen-independent proliferation in both models (Fig. 5C; Supplementary Fig. S5C; refs. 24, 25). We further tested hormone independence conferred by the AKT3 fusion using a recently described MCF7 derivative, MCF7^PIK3CAWT, in which the endogenous PIK3CA mutation in the parental line has been corrected via somatic gene targeting (28). Like parental MCF7, this line is highly estrogen-dependent for proliferation. Additionally, we directly compared effects of the AKT3 fusion in MCF7^PIK3CAWT to that of an established transforming breast cancer mutant oncogene, AKT1^E17K (29). Both AKT proteins potently induced downstream PI3K/mTORC1 signaling in MCF7^PIK3CAWT cells (Supplementary Fig. S5D). We also observed that both the AKT3 fusion and AKT1^E17K were sufficient to drive estrogen-independent proliferation in MCF7^PIK3CAWT (Fig. 5D). Numerous studies have linked such proliferation to AKT-dependent stabilization of cyclin D1 via phosphorylation of GSK3β (30). Accordingly, we observed that both the AKT3 fusion and AKT1^E17K expression triggered an increase in cyclin D1 levels in estrogen-starved MCF7^PIK3CAWT cells (Fig. 5E).

Distinct effects were observed with the RAF1 fusion, which did not confer estrogen-independent growth but instead switched the typical inhibitory effect of 4-OH tamoxifen, a competitive ER modulator, to a stimulatory one at low doses (Fig. 5F). In keeping with a receptor-dependent mechanism for this effect, the paradoxical stimulation was not observed with fulvestrant, which induces ER degradation (Fig. 5F). These findings are consistent with prior work showing that MAPK induces abnormal ER signaling that is opposed by pure estrogen antagonists but not by 4-OH tamoxifen (31). Receptor-dependent effects of each kinase fusion were further corroborated by analysis of the canonical ER transcriptional target genes CCND1 (encoding cyclin D1) and TFF1. Cells expressing each kinase exhibited enhanced basal expression of these genes, whereas we observed distinct effects of the different fusions on estrogen-induced gene expression (Supplementary Fig. S5E).

We next used an established in vivo model of HR⁺ breast cancer in order to further explore the potential oncogenic function of the AKT3 fusion (32). We transplanted T47D cells expressing either the RPS6KC1–AKT3 fusion or the control vector into the mammary fat pad of immunodeficient mice (Fig. 6A). As anticipated, control cells formed slow-growing tumors in the presence of exogenous estrogen (32). In contrast, fusion-expressing cells produced significantly larger tumors in less than one month (Fig. 6B). We then withdrew supplemental estrogen from these tumor-bearing mice, in order to mimic the effect of therapies such as aromatase inhibitors that function essentially by lowering systemic estrogen levels. As anticipated, control tumors were highly sensitive to estrogen deprivation, exhibiting significant regression in a period of days (Fig. 6C and D). Tumors expressing the AKT3 fusion, by comparison, showed no significant regression following estrogen withdrawal (Fig. 6C and D). We wished to compare these effects with those of a recognized transforming breast cancer gene and in the absence of a background PIK3CA mutation, so we repeated these experiments in MCF7^PIK3CAWT, transplanting cells expressing the AKT3 fusion, AKT1^E17K, or the control vector. Even controlling carefully for protein expression (Supplementary Fig. S5D), we observed a more highly significant increase in tumor progression with AKT3 fusion–expressing tumors than with AKT1^E17K-expressing tumors compared with the control vector (Supplementary Fig. S6A). Furthermore, upon estrogen withdrawal AKT3 fusion–expressing tumors showed minimal regression over a 2-week period. In contrast, vector-expressing MCF7^PIK3CAWT tumor regressed more than 30%, and AKT1^E17K-expressing tumors exhibited partial resistance (15% regression; Supplementary Fig. S6B and S6C).

A recent change to the standard approach for treatment of advanced HR⁺ breast cancer involves the combination of hormonal therapy with inhibitors of the cell-cycle regulatory cyclin-dependent kinases CDK4 and CDK6 (33). Accordingly, we tested whether the combination of estrogen withdrawal and the CDK4/6 inhibitor palbociclib could overcome therapeutic resistance conferred by AKT3 fusion expression in vivo (Fig. 6A). Remarkably, this combination completely abolished the advantage observed in fusion-expressing tumors, inducing a regression in both these and control tumors substantially more profound than that induced by hormone withdrawal alone (Fig. 6E). To corroborate these effects at the cellular level, we stained the resulting tumors for the proliferation marker Ki67, an established correlate of response in human breast cancer (34). Although AKT3 fusion–expressing tumors exhibited a marginally higher Ki67 fraction than control tumors following estrogen withdrawal, palbociclib treatment virtually eliminated Ki67-positive cells in both cases (Fig. 6F and G). Biochemical analysis further supported the ability of palbociclib to overcome effects of the activated AKT3 fusion. Phosphorylated retinoblastoma protein (RB1), the target of CDK4/6, was elevated in the absence of estrogen in AKT3 fusion–expressing cells compared with controls, but was completely suppressed in both cases following palbociclib treatment (Fig. 6H). Thus, this activated kinase fusion confers resistance to traditional hormonal therapy approaches that can be overcome by newly available therapeutic combinations.

Taken together, the findings above imply a contribution of identified fusions to aggressive behavior and treatment resistance in HR⁺ breast cancer. To test this possibility in patients we analyzed the impact of fusions on clinical outcomes. These included time from initial diagnosis to metastatic disease recurrence, survival from time of metastasis, and overall survival (35). To assess the potential for confounding in the association between fusion status and these outcomes, we conducted a stratified analysis comparing the distribution of clinical characteristics between fusion-positive and fusion-negative groups, controlling for the effect of cohort (Clinical Genotyping and Matched Primary/Metastasis Cohorts). Clinical characteristics did not differ significantly by fusion status, suggesting that any association between fusion status and survival would not be affected by clinical confounders (Supplementary Table S2).

Kaplan–Meier curves were then constructed to illustrate the relationship between fusion or mutation status and each clinical outcome. Cox proportional hazards models, stratified by cohort, were used to assess the association between clinical outcomes, fusions, and mutations. (Notably, as these cohorts represented patients treated prior to 2015, no patients received CDK4/6 inhibitors as either adjuvant therapy or early-line treatment in the metastatic setting.) Patients with fusion-positive tumors demonstrated a trend toward shorter time from initial diagnosis to metastatic recurrence (TTR) than those with fusion-negative tumors [Fig. 7A, hazard ratio (HR) = 1.56, P = 0.06]. By comparison, the subset of somatic oncogenic driver mutations assessed in these cohorts were not associated with differences in TTR (Fig. 7B, HR = 1.04, P = 0.83). Most strikingly, fusion-positive status was associated with highly significant decreases both in survival from time of metastatic recurrence (Fig. 7C, HR = 2.37, P = 0.01) and in overall survival (OS) from time of initial diagnosis (Fig. 7E, HR = 3.11, P = 0.0009). In contrast, the assessed somatic mutations were not significantly associated with either of these survival metrics (Fig. 7D and F).

Finally, we tested the hypothesis that as drivers of poor outcomes, fusions would be less common among patients presenting with primary (nonmetastatic) HR⁺ breast cancer, the large majority of whom are cured. Thus, we assembled a cohort of 300 such patients (Primary Diagnosis Cohort; Supplementary Table S5) and created tissue microarrays of the primary tumors, allowing us to screen the entire cohort by FISH for fusions corresponding to the most commonly implicated fusion genes. Analysis of all cases with probes for PIK3CA, ESR1, RAF1, and AKT3 (which together comprised approximately half of the fusions among our cohorts with advanced HR⁺ cancer) revealed only 12 possibly abnormal patterns by FISH. Furthermore, subsequent molecular analysis of these cases with the AMP assay demonstrated only a single bona fide fusion involving ESR1 and CCDC170 (Supplementary Table S6). Although this was not a complete analysis of all possible fusions and was by design limited to analysis of primary tumors, these results are consistent with a high prevalence of fusions selectively in patients with biologically aggressive metastatic disease and poor outcomes.

This study reveals the remarkable prevalence of expressed intergenic fusions involving key cancer genes in patients with advanced HR⁺ breast cancer. The direct contribution of these genomic alterations to the pathophysiology of these cancers is supported by a number of our observations. First, we confirmed by FISH that these genomic rearrangements are present within the tumor and are associated with protein expression. Second, we observed that the rearrangements and protein expression can be found in both primary and/or multiple matched metastases from the same patient. Third, we provide evidence that multiple novel fusions produce active and deregulated kinases that confer oncogenic phenotypes in breast epithelial models. Fourth, we demonstrate that such fusions are sufficient to disrupt hormonal-dependent proliferation, transcription, and tumor progression in HR⁺ breast cancer models both in vitro and in vivo. Notably, the most prevalent fusions detected involve ESR1, a finding that supports their relevance to the pathogenesis of HR⁺ breast cancer. Finally, we show that patients whose tumors harbor these fusions have markedly more rapid progression and shorter survival than fusion-negative patients.

Our analyses also reveal that these gene fusions appear distinct in their implications from other common somatic genetic events in HR⁺ breast cancer. For example, somatic mutations in PIK3CA are common in tumors of patients presenting with primary (curable) disease, and in this setting are generally associated with more favorable outcomes (36). In contrast, somatic mutations in ESR1 are found almost exclusively in metastatic lesions, and typically only in the setting of acquired resistance to multiple prior rounds of hormonal therapy (4, 5). The fusions we have uncovered are present in some cases in both primary and metastatic tumors, whereas in other cases only in metastases, and prior to specific therapy in the metastatic setting. Their enrichment selectively in patients who develop metastatic disease is consistent with a role in de novo treatment resistance, whereas their appearance in metastases could in those cases reflect a contribution to acquired resistance.

Our findings are in keeping with the emerging concept of gene fusions as frequent oncogenic drivers in many solid tumors (11). Until recently, however, the prevalence of these genetic events has been largely overlooked because most clinical tumor genotyping analyses (e.g., exome sequencing) do not analyze RNA and therefore fail to detect such fusions. Nonetheless, a growing body of evidence demonstrates that these species, and particularly activated kinase fusions, may represent viable therapeutic targets in multiple cancers (37). The expanding therapeutic arsenal of kinase pathway inhibitors further implies that such fusions, as well as ER gene fusions, may be immediately actionable in the setting of appropriately designed clinical trials (38). An individualized approach to targeting these fusions will likely be required, however, as evidenced by the distinct effects of different fusions on hormonal signaling and cell growth that we observe.

Despite their heterogeneity, our detailed analysis of the AKT3 fusion suggests that newly approved therapies for HR⁺ breast cancer including CDK4/6 inhibitors may overcome treatment resistance associated with some fusions. As noted, patients in this study did not receive CDK4/6 inhibitors as treatment for their primary disease, as these agents are not yet FDA approved for this indication. However, two additional patients harboring fusions we subsequently identified who developed metastatic HR⁺ disease following standard hormonal therapy did receive a combination of hormonal therapy and a CDK4/6 inhibitor for treatment of their metastatic disease. Both of these patients experienced significant clinical benefit. Because our study implies that patients with fusion-positive tumors are more likely to develop metastatic disease, the potential ability of CDK4/6 inhibitors to overcome fusion-mediated hormonal therapy resistance could portend a positive impact on relapse-free survival in the adjuvant treatment of early-stage breast cancer, an indication for which clinical trials are currently ongoing (39).

In summary, our findings reveal intergenic fusions as frequent drivers of treatment resistance and poor outcomes in advanced HR⁺ breast cancer. Given their presence in primary tumors and their association with poor survival that we demonstrate, it seems likely that routine detection of such alterations could ultimately lead to successful targeting and a major impact on long-term disease outcomes.

The Clinical Genotyping Cohort comprised a retrospective set of sequential patients with advanced ER⁺ and/or PR⁺ breast cancer who presented to Mass General and underwent local somatic genotyping by physician order between 2010 and 2014. All were tested for HER2 by IHC and/or FISH and none were positive. A single tissue sample was tested by AMP (either primary tumors or metastatic lesions depending on availability). The only exclusion was for nucleic acid samples that did not pass QC analysis for AMP testing. The Primary Diagnosis Cohort comprised a retrospective set of patients with who presented with primary ER⁺ and/or PR⁺ breast cancer and had locally available FFPE primary tumor tissue. For these two cohorts, retrospective use of discarded, clinically collected tissue, and limited clinical data was reviewed and approved by the Mass General/Harvard Cancer Center (HCC) Institutional Review Board (IRB) with waiver of consent. The Matched Primary/Metastatic Cohort included patients with available matched tissue specimens who presented with advanced ER⁺ and/or PR⁺ breast cancer to collaborating international institutions or Mass General between 2000 and 2012. All except four underwent HER2 testing by IHC and/or FISH; five were positive (all fusion-negative) and none received anti-HER2 therapy. Eligible patients signed written informed consent for participation in the study, which was approved by the HCC IRB and the individual local IRBs, all of which follow the U.S. Common Rule or the International Ethical Guidelines for Biomedical Research Involving Human Subjects (CIOMS). Concordance between fusion detection in primary and metastatic tumors for this cohort is illustrated in Supplementary Fig. S1A.

The methodology for each of these has been described previously (12, 15). Briefly, total nucleic acids containing total RNA and genomic DNA were extracted from FFPE tissue specimens, using the Agencourt FormaPure Kit for FFPE Tissue (Beckman Coulter). For AMP analysis, first- and second-strand complementary DNA (cDNA) synthesis was performed using a combination of SuperScript III (Life Technologies), DNA Polymerase I (Enzymatics) and RNAse H (Enzymatics). The resulting cDNA/DNA underwent end repair (End-Repair Mix, Enzymatics), adenylation (Klenow Exo-, Enzymatics; Taq Polymerase, Life Technologies), and ligation (T4 DNA Ligase, Enzymatics) with a universal Y adaptor containing a sample barcode, universal priming site, and Illumina sequencing priming site. Ligated libraries were subjected to two rounds of nested PCR at 10 to 14 cycles each for target enrichment (Platinum Taq Polymerase, Life Technologies). The first round of PCR was performed using a primer complementary to the universal adapter and a first pool of target-specific primers (Operon). A second round of PCR is executed using a 3′ nested universal adapter primer downstream of the first adapter primer and a second pool of 3′ nested target-specific primers downstream of the respective initial first-pool target primers. These nested primers are each 5′ tagged with a common sequencing adapter which, in combination with the first half-functional universal adapter, creates target amplicons ready for clonal amplification and sequencing. Libraries are quantitated using quantitative PCR (Kapa Biosystems), normalized, and processed for sequencing on an Illumina Nextseq 500 instrument according to the manufacturers' standard protocol. Sequencing reads were demultiplexed using Illumina's bcl2fastq script. BWA-MEM was used to align reads to their respective positions according to the NCBI reference genome hg19. Marking of duplicate reads and base quality recalibration is performed using Picard tools and GATK. Suspect chimeric fusion reads after alignment are then called and annotated with a custom script as described (12). For the cohort an average coverage of 200× was obtained per sample, and fusions were prioritized by filtering out known recurrent sequencing artifacts, identifying at least 5 supporting reads, and confirming in-frame status of predicted fusion transcript translated products. The gene and exon targets for fusion analysis are listed in Supplementary Table S1. Site-specific somatic genotyping was performed as described (15), using an adaptation of the Applied Biosystems (ABI) Prism SNaPshot Multiplex system originally developed to detect single nucleotide polymorphisms (SNP). In essence, multiplex PCR with gene-specific primers was followed by primer extension with differentially fluorescent-labeled nucleotides, then extension products were resolved by an automatic capillary sequencer (ABI PRISM 3730 DNA Analyser, Life Technologies/Applied Biosystems) as described (15). Data analysis for SNaPshot genotyping was performed with GeneMapper Analysis Software version 4.0 (Life Technologies/Applied Biosystems). The gene and residue targets for fusion analysis are listed in Supplementary Table S4.

Gene rearrangements were analyzed using a break-apart probe strategy, with two probes spanning the 3′ (spectrum orange label) and 5′ (spectrum green label) ends. In this approach, a fused signal is observed in normal cases, and split signals, green and orange, are seen when translocation is present. BAC clone (probe) searches were performed using the University of California Santa Cruz (UCSC) Genome Browser (http://genome.ucsc.edu/) mapped to December 2013 (GRCh38/hg38) Human Genome Assemblies. BACs were purchased from Children's Hospital Oakland Research Institute (CHORI, Oakland, CA; http://bacpac.chori.org/). BAC clones that showed nonspecific hybridization were discarded. All BAC clones used are shown in Supplementary Fig. S1C. Purified BAC DNA was amplified and directly labeled with spectrum orange or spectrum green dUTPs (Enzo Life Sciences) by nick translation (Abbott Molecular Inc.) according to the manufacturer's protocol. Tissue sections (5 μm) were deparaffinized, pretreated with Tris/EDTA, pH 7.0, and treated with Digest-all (Thermo Fisher) at 37°C for 90 minutes. After washing with 2 × saline sodium citrate (SSC), sections were dehydrated in ethanol and dried. Slides were co-denatured with FISH probes using a Hybrite slide processor (Abbott Molecular Inc.) at 85°C for 5 minutes. Hybridization was performed overnight at 40°C. After hybridization completion, slides were washed two times in 2 × SSC/0.1% Nonidet-40 for 3 minutes each at 72°C, rinsed in water and mounted in Vectashield (Vector Laboratories) containing 4,6-diamino-2-phenylindole (DAPI) as a counterstain. The probes were visualized using an Olympus BX61 microscope and analyzed with Cytovision software (Leica Biosystems). One hundred nuclei were scored in each TMA core, and the positive cases were subsequently confirmed in their tissue blocks. In the tissue blocks, tumor area was located before analysis. The number of “break-apart” or split signals was searched, and a separation of at least two signal diameters was considered as the threshold for a positive split signal. The presence of 5% or more of split signals was indicative of true positive result.

Cells were maintained at 37°C in 5% CO₂. MCF7 (ATCC) and T47D (ATCC) cells were grown in RPMI (Lonza) supplemented with 10% FBS (SAFC), 1% penicillin (Gibco), streptomycin (Gibco), and glutamine (Gibco). MCF10A cells were grown in DMEMF12 (Lonza), 5% horse serum (Gibco), 20 ng/mL EGF (Sigma), 10 μg/mL insulin (Sigma), 1 ng/mL cholera toxin (Sigma), and 100 μg/mL hydrocortisone (Sigma). All cell lines except MCF7^PIK3CAWT were obtained from the MGH Center for Molecular Therapeutics cell bank in 2016 and 2017 and underwent high-density SNP typing to confirm their identity. For MCF7^PIK3CAWT we verified the genotype and cell line identity by a specific PCR assay at the time of receipt (2017) as described (28). All experiments shown were performed within less than 6 months' passage of all lines since acquisition.

To study the fusion-driven effects on canonical E2 targets gene expression, MCF7 cells were cultured in phenol red-free medium with charcoal–dextran-stripped fetal bovine serum for 72 hours, and stimulated with 100 pmol/L of 17β-estradiol (Sigma) for 12 hours.

For 4-OH tamoxifen (selleckchem) and fulvestrant (selleckchem) sensitivity studies, cells were maintained in phenol red-free medium with charcoal–dextran-stripped fetal bovine serum for 48 hours, and 1,000 cells per well were seeded in 96-well plates and exposed to varying doses of 4-OH tamoxifen (10 pmol/L to 1 μmol/L) and fulvestrant (1 nmol/L to 1 μmol/L). Cell proliferation was assayed after 4 days for fulvestrant and 6 days for 4-OH tamoxifen. For MK-2206 (selleckchem) and trametinib (selleckchem) sensitivity studies, cells were cultured in DMEM with 10% FBS, and 1,000 cells were seeded in 96-well plates and exposed to varying doses of MK2206 (1 nmol/L to 1 μmol/L) and trametinib (100 pmol/L to 100 nmol/L) and assayed after 3 days. To determine cell proliferation upon drug treatment, CellTiter-Glo Luminescent Cell Viability Assay (Promega) was used according to the manufacturer's recommendations.

The cDNA fragments for CTNNBL1-RAF1, ACTL6A-PIK3CA, and RPS6KC1-AKT3 were amplified from pDONR223-RAF1, pDONR223-AKT3, and pDONR223-RPS6KC1 (Addgene; ref. 40) and pCIG PIK3CA wild-type (Addgene; ref. 41), respectively. CTNNBL1 was amplified from T47D cells, and an oligonucleotide was designed for ACTL6A (primers and oligos listed in Supplementary Table S7). Constructs were generated by using pENTR Direction TOPO cloning kit (Invitrogen). To generate lentiviral constructs, pLenti CMV Puro DEST (w118-1; Addgene; ref. 42) was used. After verifying by sequencing, pLenti-based kinase fusions were cotransfected into HEK293T cells with lentiviral packaging plasmids pMD2.G and psPAX2 (Addgene) by using the CalPhos Mammalian Transfection Kit (Clontech Laboratories) according to the manufacturer's instruction. The conditioned media containing lentiviral particles were collected 36 hours after transfection and filtered using 0.45-μm pore filter (Millipore). The filtered media were then used to infect target cells. Polybrene (Sigma) was added into filtered media at a final concentration of 10 μg/mL to increase the infection efficiency. The infected cells were selected with 1 μg/μL puromycin (Sigma) 72 hours after infection. Expression of fusion constructs was confirmed via Western blotting.

Total RNA was isolated as described previously (43), and qPCR primers are listed in Supplementary Table S7.

For total protein extraction, cells were lysed in RIPA buffer [10 mmol/L Tris–HCl pH 7.5, 150 mmol/L NaCl, 1 mmol/L EDTA, 1% (w/v) sodium deoxycholate, 0.1% (w/v) SDS, 1% (v/v) NP40, proteinase inhibitor cocktail, phosphatase inhibitor cocktail] for 30 minutes at 4°C.

The protein samples were mixed with SDS sample buffer and boiled for 10 minutes before being subject to SDS-PAGE. The protein samples on the SDS-PAGE gel were then transferred onto PVDF membrane (Milipore), which was blocked by 5% nonfat milk in PBST (PBS plus 0.02% Tween 20) at room temperature for 1 hour. Then, the PVDF membrane was incubated with primary antibodies diluted in 3% BSA in PBST at 4°C overnight and horseradish peroxidase-conjugated secondary antibodies (Sigma) diluted in 3% BSA in PBST at room temperature for 2 hours. The signal was detected by enhanced chemiluminescence solution (PerkinElmer).

Antibodies used are listed in Supplementary Table S8.

Three-dimensional culture of MCF10A was performed as described previously (44). Briefly, fusion-expressing MCF10A cells were diluted to a final concentration of 25,000 cells/mL, and Matrigel (Corning) was added in a 1:1 ratio. Matrigel–cell mixture was plated in 8-well chamber slides (lab-tek), 400 μL in each well. Cells were grown in a 5% CO₂ humidified incubator. MCF10A acini cultured in Matrigel were fixed by using an immunofluorescence application solution kit (Cell Signaling Technology) according to the manufacturer's recommendations. To stain cell–cell junctions, acini were stained with anti–β-catenin (Supplementary Table S8) at 1:100 dilution and secondary Alexa Fluor 488 at 1:200 dilution. To distinguish nuclei, Hoechst 33342 (molecular probes) was used. Prolong Antifade Reagent (Molecular Probes) was used to preserve the slides. Confocal microscopy (Zeiss LSM510 confocal microscope LSM510) was performed as described in Debnath and colleagues (44) with pinhole size of 0.7 to 0.9 μm. Imaris software (bitplane) was utilized for quantification, and at least 200 acini were analyzed per construct. For analysis of cell death during 3-D culture, ethidium bromide (Sigma) was used. After 8 days, 3-D cultures were incubated for 15 minutes with 1 μmol/L ethidium bromide (Sigma) in PBS. Ethidium bromide–positive cells were detected immediately after incubation on the rhodamine channel. Two hundred structures per construct were counted for quantification (20).

All animals were housed and treated in accordance with protocols approved by the Subcommittee on Research Animal Care at the Massachusetts General Hospital. Female athymic nu/nu mice (5–6 weeks) were anesthetized with Avertin (2.5% 2,2,2 tribromoethanol in sterile 1 × PBS) at 16 μL/g (45). Animals were implanted with 60-day slow-release pellets containing 0.72 mg of 17β-estradiol (Innovative Research of America; ref. 46). Xenograft tumors were generated by subcutaneous injections of 8 × 10⁶ T47D cells expressing RPS6KC1-AKT3 or empty vector suspended in 1:1 Matrigel into the mammary fat pad of nude mice. Tumors were measured twice weekly by caliper, and volumes were calculated by the formula 0.5(length) × (width)². To mimic tumor growth under E2 deprivation conditions, 17b-estradiol pellets were removed. Palbociclib (LC Labs) was given daily by oral gavage at 150 mg/kg/day for 7 days. Tumor growth was monitored as indicated twice per week or every 2 days (for palbociclib) and tumor volume was measured using the formula ½(length × width²).

We evaluated the effect of fusions and somatic mutations on three clinical outcomes: overall survival, time from initial diagnosis to metastasis, and survival from time of metastasis. The analyses included data from 101 patients in the Clinical Genotyping Cohort (9 of 110 patients tested were excluded due to missing date of primary diagnosis) and 80 in the Matched Primary/Metastasis Cohort (1 of 81 excluded due to missing data), including 21 who had metastatic disease at the time of initial diagnosis. All patients underwent both AMP and SNaPshot genotyping analysis with the exception of 5 patients lacking SNaPshot data in the Clinical Genotyping Cohort, and 13 lacking SNaPshot data and 18 lacking AMP data in the Matched Primary/Metastasis Cohort. Overall survival was defined as the time from initial diagnosis to death (event) or last follow-up (censored), and the analysis of overall survival included patients with metastatic disease at the initial diagnosis. The analysis of time from initial diagnosis to metastasis excluded patients who were metastatic at initial diagnosis, and event times were observed for all included patients (no censoring). The analysis of time from metastasis to death included both patients who were and were not metastatic at diagnosis. Patients with inadequate data to define a given event time were excluded from the analysis of that event, and those without data for fusion (or mutation) status were excluded from the analyses of the effect of fusion (or mutation) status on each clinical outcome. For the analyses of overall survival and survival from time of metastasis, patients who were known to be deceased but had unknown death dates were treated as alive/censored at the date of last follow-up. Kaplan–Meier curves were constructed to illustrate the relationship between fusions and mutations and each clinical outcome. Cox proportional hazards models were used to assess the association between clinical outcomes and fusions and mutations. These models included the study cohort as a stratification factor which allows for a different underlying hazard function for the two cohorts. Hazard ratios estimated from these models reflect the increase in the hazard rate for a given event associated with positive status for fusion or mutation, with an associated two-sided P value of less than 0.05 indicating a statistically significant association. All other experimental results were analyzed using unpaired two-tailed Student t test or ANOVA, as indicated in figure legends. Groups that were statistically compared shared a similar variance, as shown in the figures. P < 0.05 was considered statistically significant.

Disclosure of Potential Conflicts of Interest

Z. Zheng is the inventor of the patented AMP method and is an ArcherDX equity holder. D.M. Finkelstein is a consultant/advisory board member for Gilead and Janssen. L.P. Le has ownership interest (including patents) in ArcherDx and is a consultant/advisory board member for the same. A.J. Iafrate has ownership interest (including patents) in ArcherDx and is a consultant/advisory board member for Roche and Chugai. No potential conflicts of interest were disclosed by the other authors.

Conception and design: K.J. Matissek, M.L. Onozato, S. Sun, Z. Zheng, A. Schultz, K. Patel, S.V. Saladi, M. Debiasi, G. Werutsky, A. Bardia, P.E. Goss, D.C. Sgroi, A.J. Iafrate, L.W. Ellisen

Development of methodology: K.J. Matissek, M.L. Onozato, S. Sun, Z. Zheng, A. Schultz, G. Werutsky, L.P. Le, A.J. Iafrate

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): K.J. Matissek, M.L. Onozato, S. Sun, Z. Zheng, K. Patel, P.-L. Jerevall, S.V. Saladi, M. Tavallai, T. Badovinac-Crnjevic, C. Barrios, N. Beşe, A. Chan, Y. Chavarri-Guerra, E. Demirdögen, H. Gomez, P. Liedke, G. Werutsky, A. Bardia, P.E. Goss, D.C. Sgroi, A.J. Iafrate

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): K.J. Matissek, M.L. Onozato, S. Sun, Z. Zheng, A. Schultz, J. Lee, K. Patel, P.-L. Jerevall, S.V. Saladi, A. Macleay, M. Tavallai, A. Chan, N. Horick, D.M. Finkelstein, L.P. Le, A. Bardia, D.C. Sgroi, A.J. Iafrate, L.W. Ellisen

Writing, review, and/or revision of the manuscript: K.J. Matissek, M.L. Onozato, Z. Zheng, P.-L. Jerevall, T. Badovinac-Crnjevic, C. Barrios, A. Chan, Y. Chavarri-Guerra, M. Debiasi, Ü. Egeli, H. Gomez, P. Liedke, G. Werutsky, J. St. Louis, N. Horick, A. Bardia, P.E. Goss, D.C. Sgroi, A.J. Iafrate, L.W. Ellisen

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): K.J. Matissek, A. Schultz, J. Lee, A. Chan, Ü. Egeli, S. Gökgöz, I. Tasdelen, S. Tolunay, J. St. Louis, P.E. Goss, A.J. Iafrate

Study supervision: G. Werutsky, L.W. Ellisen

Other (provided all clinical data and the majority of the biobank): P.E. Goss

This work was supported in part by the Avon Foundation Breast Cancer Crusade and the Tracey Davis Memorial Fund. We gratefully acknowledge international collaborators who provided matched primary/metastatic breast cancer specimens, from the Latin America Cooperative Oncology Group (LACOG) including Laura Voelcker, Carlos Sampaio (Pontificia Universidade Catolica do Rio Grande do Sul), Giuliano Borges (Clinica de Neoplasias Litoral), Gustavo Ismael (Fundaçao Amaral Carvalho), Yeni Neron (CEPON), Gabriel Prolla (Hospital Mae de Deus), and Susanne Crocamo (INCA). We also thank Josh Lauring (Johns Hopkins University) for the generous gift of MCF7^PIK3CAWT cells. D.C. Sgroi was funded in part by a grant from the Breast Cancer Research Foundation.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.