Genomic landscape of estrogen receptor-positive HER2-negative advanced breast cancer with acquired resistance to aromatase inhibitors: Identification of an ESR1 alteration-related gene signature

Abstract

Background

Treatment of advanced estrogen receptor-positive HER-2-negative breast cancer is based on hormonal therapy with aromatase inhibitors for postmenopausal women. However, acquired endocrine resistance is unavoidable at some point in the advanced or metastatic stage, and its underlying molecular mechanisms remain to be fully elucidated. The study prospectively included patients with advanced or metastatic breast cancer who had relapsed or progressed following treatment with a non-steroidal aromatase inhibitor (AI) and were treated with exemestane (a steroidal AI) and everolimus. The objective was to perform DNA and RNA analyses in order to gain a deeper understanding of the genomic and transcriptomic landscape of endocrine-resistant advanced BC.

Material and methods

We selected 65 patients included between 2015 and 2018 in the SAFIRTOR trial (NCT02444390). NGS-based gene panel of 65 genes (SAFIR02 core panel), Comparative genomic hybridization array and RNA-sequencing assessed single nucleotide variations, copy-number variations and gene expression and fusion genes of interest, respectively.

Results

The most prevalent genomic alteration was ESR1, observed in 49% (32/65) of cases, with most being activating missense mutations. Two cases of ESR1 fusions were identified. The observed alterations were not mutually exclusive with those from NF1 or ERBB2. Differential expression analysis in the presence or absence of ESR1 alterations showed significant enrichment of the ESR-mediated signaling pathway in tumors with ESR1 alterations (p < 0.005). 17 genes were identified as significantly differentially expressed, forming a transcriptional signature with 81.3% sensitivity and 84.3% specificity (p < 0.0001). The oxidative phosphorylation (OXPHOS) pathway was significantly associated with resistance to everolimus (p = 5.05 × 10^-10). 16 genes were differentially expressed between responders and non-responders, forming a OXPHOS gene signature distinguishing two groups with markedly different outcomes.

Introduction

Estrogen receptor-positive HER2-negative (ER + /HER2-) represents the most common form of breast cancer (BC), with an estimated age-adjusted rate of 87.2 new cases per 100,000 women in 2023 worldwide [1]. In the recurrent or metastatic setting, the strategy of combining aromatase inhibitors (AI) with cyclin-dependent kinase 4 and 6 (CDK4/6) inhibitors has demonstrated efficacy as a first-line therapeutic approach [2–5]. Although most patients initially benefit from this strategy, they eventually develop endocrine resistance, resulting in an incurable disease [6].

Lately, interest is growing for the molecular biology of advanced endocrine-resistant breast cancer, as cancer molecular features evolve over the time, especially with drug exposure [7]. Enhanced understanding of these molecular mechanisms not only helps to understand disease biology but also yields prognostic and predictive biomarkers, thereby facilitating the development of targeted therapies. Endogenous estrogens, such as estradiol, exert their biological effects through binding to estrogen receptor-α (ER-α), a nuclear receptor belonging to the steroid receptor family [8]. ER-α functions as a transcription factor that regulates the expression of a broad range of target genes by binding to specific DNA sequences known as estrogen response elements (EREs). Upon ligand binding, estradiol induces a conformational change in ER-α, promoting its dimerization - typically forming homodimers -, which then translocate into the nucleus. There, ER-α directly interacts with EREs in the promoter or enhancer regions of target genes, recruiting coactivators and other transcriptional machinery to modulate gene expression [9]. This transcriptional regulation influences a variety of physiological processes, including cell growth, differentiation, and reproductive functions. Structurally, ER-α contains a ligand-binding domain (LBD), responsible for estrogen binding, and a DNA-binding domain (DBD), which mediates its interaction with EREs to control gene transcription [10]. Physiologically, most BC cases are ER-α positive at diagnosis [11]. Endogenous estrogen mediates its effects by binding to ER-α, leading to expression of genes controlling proliferation and cell survival. However, a large proportion of patients will encounter relapse with de novo or acquired resistance disease. The most described AI resistance mechanism is the presence of an estrogen receptor 1 (ESR1) alteration, mostly acquired somatic mutations in the carboxy-terminal ligand-binding domain of the protein causing ligand-independent activation [12], but also ESR1 translocations or ESR1 amplifications [6], whereas ESR1 alterations are quite rare (less than 3%) in primary tumors [13]. These alterations lead to aberrant cross-talk between ESR1 and growth factor signaling pathways, as well as alterations in the balance of coactivators and corepressors [6, 14, 15]. These ESR1 mutations account for only about half of the molecular mechanisms of acquired resistance to AI in combination [7].

Phosphatidylinositol 3-kinase (PI3K)/protein kinase B (AKT)/mammalian target of rapamycine (mTOR) pathway is also implicated in mediating treatment resistance [16, 17]. Mutations in the α-catalytic subunit of PI3K (PIK3CA; which encodes p110α) are the most frequent genetic abnormality in luminal-type breast cancer (occurring at a frequency of 30–40%) [18]. Hence, everolimus, an mTOR inhibitor, is a standard treatment in combination with exemestane after non steroidal AI and CDK 4/6 inhibitors in advanced breast cancer [19]. We recently showed that the activation of the mTOR pathway, highlighted by phospho-4EBP1 (p4EBP1) staining on tissue might be associated with enhanced efficacy of everolimus [20].

Additional acquired resistance mechanisms, such as the emergence of ERBB2 mutations or NF1 loss-of-function clones, have been suggested [7]. Whether these additional mutations exhibit mutual exclusion and their potential for targeted therapy remain yet to be determined [7, 16].

In order to better delineate the molecular mechanisms underlying acquired resistance to AI in advanced breast cancer, we conducted an ancillary study to the SAFIRTOR trial [20]. We performed next-generation sequencing (NGS), comparative genomic hybridization array, and RNA-sequencing analysis on tissue biopsy. Through this comprehensive approach, we aim to decipher both the genome and transcriptome landscape of endocrine-resistant advanced BC.

Material and methods

Patients and study design

The cohort comprised 65 patients included in SAFIRTOR, a single-arm prospective biomarker study sponsored by UNICANCER (NCT02444390). Key inclusion criteria were as follows: postmenopausal women aged ≥ 18 years with histologically-proven ER-positive/HER2-negative locally advanced disease not amenable to curative surgery or radiotherapy or metastatic BC, undergoing second-line endocrine therapy or more after recurrence or progression with a non-steroidal AI in the adjuvant or metastatic setting. Patients had to be experiencing progressive disease under endocrine therapy at the time of inclusion. Other inclusion criteria included measurable or evaluable disease, a metastasis at a site accessible for a biopsy, except bone metastases, and WHO Performance Status 0 or 1. Exclusion criteria included contraindications to everolimus plus exemestane treatment, prior treatment with mTOR inhibitors, more than one line of chemotherapy in the metastatic setting, symptomatic visceral disease, spinal cord compression and/or symptomatic or progressive brain metastases, hematopoietic function or organ impairment, and contraindications to biopsy. All patient were treated with the exemestane-everolimus combination. We considered as responders to everolimus patients encompassing complete, partial or stable response after 6 months of everolimus plus exemestane and non-responders to everolimus the other patients (35 vs. 30 cases, respectively).

Ethics approval and consent to participate

All patients provided written informed consent. All methods were performed in accordance with the relevant guidelines and regulations. The study was approved by ANSM and the national ethics committee.

Procedures and assessments

Patients underwent biopsy of the metastatic site, or the primary tumor in case of de novo metastatic disease, for the collection of at least three core biopsies. In case of multiple metastatic sites, the choice of the metastatic site to be biopsied was left to the investigator’s discretion. Following biopsy, patients started treatment with everolimus and exemestane as per approved indications.

Determination of genomic alterations in SAFIR-TOR trial

Patients across France had access to molecular testing in five regional molecular cancer genetics platforms labeled by UNICANCER for the SAFIR02 trial (NCT02299999). Only samples with a pathologist-assessed tumor cell fraction of more than 30% were selected for DNA extraction and genomic testing. Copy-number variations (CNV) were detected using comparative genomic hybridization array (CGHa). Single nucleotide variant (SNV) and short (<50 bp) insertion/deletion (indels) were detected with an NGS-based gene panel. Copy-number variations (CNV) were detected using comparative genomic hybridization array (CGHa) using the OncoScan FFPE Assays Kits (Affymetrix, Thermo Fisher Scientific). Single nucleotide variant (SNV) and short (<50 bp) insertion/deletion (indels) were detected with an NGS-based gene panel. DNA (10 ng) was used to perform the initial PCR step (17 cycles). Amplicons were then partially digested using the enzyme FuPA to eliminate extremities corresponding to primer sequences. The digested product was ligated with adaptors and barcodes, then amplified and purified. Adaptors contained specific indices (barcodes) different for each sample so that libraries from different samples could be combined before sequencing. Quality and quantity assessment of DNA libraries was achieved using a Qubit 2.0 Fluorometer (Thermo Fisher Scientific) and/or BioAnalyzer profiling. After equimolar pooling of libraries, the final solution was sequencing using either a MiSeq instrument (Illumina), an Ion Torrent PGM or an Ion Torrent GeneStudio S5 (Thermo Fisher Scientific), depending on the various regional molecular cancer genetics platforms. The gene panel encompasses 65 genes (SAFIR02 core panel, Supplementary Table 1) [21].

Bioinformatics analysis and biological interpretation

Focal amplifications were defined as segments of less than 10 Mb with log2 ratio > 1.58 (equivalent to a copy number of 6 or more) and homozygous deletions as segments of less than 10 Mb with log2 ratio < -1. CNV calls were validated by visual inspection of copy-number profiles.

The pathogenicity of small variants and SNV was determined by automated processing by the OncoKB annotator tool [22] and manual curation using the literature and authoritative databases (COSMIC [23], CancerHotspot [24], OncoKB [25]). Were classified as pathogenic: all variants labeled as oncogenic, likely oncogenic, or predicted oncogenic by the OncoKB-annotator tool ; all truncating variants in tumor suppressor genes; and variants with sufficient evidence for their oncogenic effect in the literature and databases in oncogenes and tumor suppressor genes (e.g., hotspot missense variants).

Transcriptomic analysis

Isolation of RNA from frozen core biopsies was performed using the AllPrep RNA Mini kit (Qiagen, Hilden, Germany) according to the manufacturer’s protocol. RNA-seq was performed using the TruSeq RNA exome. Raw sequencing data was processed with STAR [26] for alignment and expression level quantification. Fusion gene were identified by three detection tools: Arriba (version 1.2.0) [27], FusionCatcher (version 1.20) [28], and StarFusion (version 1.8.1) [29], validated with FusionInspector version 2.2.1 [30], and integrated by a meta calling tool, FusionReport (version 2.0.1) [31]. Only known pathogenic gene fusions, reported in breast cancers or other cancer types, were retained.

Downstream data analyses were performed using R version 4.1.2 [32]. The EdgeR and limma packages were used for differential expression analysis. Overrepresentation Analysis (OA) and gene Set Enrichment Analysis (GSEA) were conducted using the clusterProfiler R package [33] to query the Reactome Pathway Knowledgebase (REAC) [34] and the KEGG pathway database [35]. Pathway enrichment was considered statistically significant at an adjusted p-value threshold of < 0.05 (Fisher’s exact test adjusted with Benjamini-Hochberg method).

Hierarchical clustering was conducted using the hclust function from the stats package (v4.3.0), applying the “complete linkage” method with Euclidean distance as the measure of dissimilarity on transcripts per million (TPM) counts produced by STAR.

Statistical analysis

Statistical analyses were performed using R version 4.1.2. Numerical variables were expressed as the median or mean and standard deviation, as appropriate. Categorical variables were expressed as the count and percentage. Variables were tested using Fisher’s exact test, the chi-square test, or the Mann-Whitney U-test, as appropriate. PFS was calculated from the date of treatment initiation to the date of event defined as the first documented progression or death from any cause. Patients with no event at the time of analysis were censored at the date of last tumor assessment. OS was calculated from the date of treatment initiation to the date of death from any cause. Patients who were alive at the time of analysis were censored at the date of last contact. PFS and OS outcomes were compared between subgroups using log-rank tests. Survival analyses were performed using the R package Survival (v3.7) on clinical and molecular features, employing the Kaplan-Meier method. Statistical significance threshold was set at 0.05.

Results

Patient characteristics

Between 24 April 2015 and 15 March 2018, 165 patients were screened and 150 eligible patients were enrolled in the SAFIRTOR study. Targeted next-generation sequencing, CGH array and RNA sequencing were available for 65 patients who were included in this ancillary study. Patient’s characteristics are shown in Supplementary Table 2.

Genomic landscape of AI-resistant breast cancer

We report here the altered genes (with frequency ≥ 2%) among 65 BC patients who progressed or relapsed after AI, using targeted NGS for SNV, CGH array for CNV and RNA sequencing analysis for gene fusion detection (Fig. 1). ESR1 alterations were detected in 49% of these patients (32/65), mainly activating missense mutations but also 2 cases of ESR1 fusions (Supplementary Table 3). In one of these two cases, two distinct gene fusions were identified, involving different partners: IMPG1 located 75 Mb upstream of ESR1 on the short arm of chromosome 6 (intra-chromosomal rearrangement), and TDRD1 on chromosome 10 (inter-chromosomal rearrangement). Most ESR1 mutations were located in the Ligand-Binding Domain (LBD), as shown in Supplementary Fig. 1. 3 patients were carrying minor hotspot ESR1 mutations. We identified 37% patients (24/65) with alterations in PIK3CA, mainly single activating missense mutations, but also four cases of double mutations of the PIK3CA gene, as previously reported [19]. Focal amplifications were detected in 23% of patients (15/65) for MYC, 20% (13/65) for CCND1, and 18% (12/65) for FGFR1.We identified alterations in TP53 in 15% of patients (10/65). Additionally, we found driver missense mutations in ERBB2 in 6% (4/65) of patients and inactivating alterations in NF1, including frameshift indels, nonsense mutations, and fusions, in 6% (6/65) of patients. These different alterations were not statistically mutually exclusive within our cohort (Fisher’s exact test). There were no statistically significant differences in the distribution of these alterations between responders and non-responders to everolimus (Supplementary Fig. 2, Supplementary Table 4).

Fig. 1. Molecular landscape of 65 breast cancer with acquired resistance to aromatase inhibitors.

Each column represents a patient, and each row corresponds to a specific gene. The proportion of patients with alterations in key genes is indicated on the right. Only genes found altered in more than 2% of samples are represented.

Differential expression gene analysis between ESR1 mutated/fusions tumors and ESR1 wild-type tumors

Given that ESR1 is the most frequently acquired mutation under selective pressure from aromatase inhibitors, we investigated the differences in the expression profiles of tumors based on the presence or absence of an ESR1 pathogenic alteration. We conducted differential expression analysis to compare patients with ESR1 mutations or fusions (n = 32) and those without (n = 33), aiming to identify an ESR1 alteration-related transcriptional signature.

We identified a set of 494 differentially expressed genes (Supplementary Table 5). OA in the REAC database revealed, as expected, significant enrichment of the ESR-mediated signaling pathway (REAC identifier R-HSA-8939211), with 17 differentially expressed genes, and the Estrogen-dependent gene expression pathway (R-HSA-9018519), with 15 differentially expressed genes, all of which are included in the 17 genes from the ESR-mediated signaling pathway. These genes are IGF1R, GREB1, FKBP4, PGR, NRIP1, MYC, HSP0AB1, ESR1, CXXC5, H2AFJ, H2BC21, HSPB1, USF2, H2BC5, H4-16, KCTD6, and NCOA1 (Supplementary Fig. 3, Supplementary Table 6), all but NCOA1 (a transcriptional coactivator for nuclear hormone receptors) overexpressed in ESR1-altered tumors. GSEA also demonstrated significant enrichment of the ESR-mediated signaling pathway (normalized enrichement score (NES) = 2.2, adjusted p-value = 7.73 × 10⁻¹⁰; Fisher’s exact test adjusted with Benjamini-Hochberg method) (Supplementary Fig. 4A). Hierarchical clustering based on the expression levels of these 17 genes identified two groups (A and B) with markedly different ESR1 alteration frequency (Fig. 2), predicting with a sensitivity of 81.3% and a specificity of 84.3% the presence of an ESR1 alteration (p-value < 0.0001, chi-squared). It is worth noting that the 2 ESR1 fusions are clustered into the ESR1-altered group (B). Regarding the minor ESR1 mutations: one p.Glu380Gln mutation as well as the p.Ser463Pro mutation were located within ESR1-altered group (B), while the other p.Glu380Gln mutation was located within ESR1-wild type group (A). No KEGG database pathway was found to be significantly enriched by OA.

Fig. 2. Heatmap depicting supervised clustering based on the 17-gene signature, enabling the prediction of the presence or absence of an ESR1 alteration (mutation or fusion).

The color chart on the right side represents the level of expression of the 17-gene signature. Red color represents a high level of expression while the blue color represents a poor level of expression. Two distinct clusters emerge: Group A (left), where most ESR1 wild-type tumors are clustered, and Group B (right), which contains the majority of ESR1-altered tumors.

Gene expression is mediated by ER/OXPHOS pathway in responders to everolimus

We then examined the transcriptomic profiles of patients who responded to everolimus in our cohort, identifying 211 differentially expressed genes compared to non-responders (Supplementary Table 7). ORA against the KEGG database revealed significant enrichment of 14 substantially overlapping pathways (Supplementary Table 8). The oxidative phosphorylation pathway (OXPHOS, KEGG identifier hsa00190) showed the highest statistical significance of enrichment, (p-value = 6.37 × 10⁻¹⁰, Fisher’s exact test adjusted with Benjamini-Hochberg method) and 16 genes from this pathway were differentially expressed between responders and non-responders. GSEA also demonstrated significant enrichment of the OXPHOS pathway (NES = 1.9, adjusted p-value = 2.51 × 10⁻⁵, Fisher’s exact test adjusted with Benjamini-Hochberg method, Supplementary Fig. 4B). ORA on the REAC database yielded largely similar results, with significant enrichment of mitochondrial respiration-related pathways (Supplementary Table 8).

Hierarchical clustering on the expression levels of the 16 OXPHOS genes was able to identify 2 groups with distinct outcomes under everolimus treatment: an OXPHOS-low expression group (A) with a 71,4% (19/27) chance of response, and an OXPHOS-high expression group (B) with a 42,1% (16/38) chance of response (p = 0.024, Chi-squared test) (Fig. 3). Survival analysis in the whole cohort (Fig. 4a) revealed a significant difference in overall survival between the groups (p-value = 0.041, log-rank test), with a median survival time of 38.1 months in the OXPHOS-low compared to 29.4 months in the OXPHOS-high group. There was no statistically significant association between ESR1 status and OXPHOS expression group (p-value = 1, Chi-squared test). In consequence, we then analyzed overall survival in ESR1 wild-type patients (Fig. 4b) and ESR1-altered patients (Fig. 4c) according to OXPHOS expression levels. In ESR1 wild-type patients, no significant difference in overall survival was observed between the OXPHOS-low and OXPHOS-high groups (p = 0.35, Log-Rank test). In contrast, among ESR1-altered patients, overall survival was prolonged in the OXPHOS-low group compared to the OXPHOS-high group, with a p-value of 0.064, approaching statistical significance (Log-Rank test). Progression-free survival (PFS) analysis in the whole cohort did not reveal a significant difference between the groups (p-value = 0.66, log-rank test), with a median progression-free survival time of 10.1 months in OXPHOS-low compared to 6.6 months in OXPHOS-high group (Supplementary Fig. 5A) as well as in ESR1 wild-type tumors, while in ESR1-altered it seemed that PFS was higher in OXPHOS-low versus OXPHOS-high (Supplementary Fig. 5B and 5C, respectively).

Fig. 3. Heatmap representing the supervised clustering performed based on the 16-gene signature of the OXPHOS pathway, and its ability to predict resistance to everolimus.

The color chart on the right side represents the level of expression of the 16-genes signature. Red color represents a high level of expression while the blue color represents a poor level of expression. The Kaplan-Meier plot was restricted to 55 months of follow-up due to the small size of the at-risk size group beyond this period.

Fig. 4. Overall survival according to OXPHOS genes signature expression levels.

a Overall survival in the whole cohort according to the 16 OXPHOS genes signature: group 1 (low expression of OXPHOS genes, OXPHOS-low), and group 2 (high expression of OXPHOS genes, OXPHOS-high), Log-Rank test. b Overall survival in the ESR1 wild-type tumors according to the 16 OXPHOS genes signature: group 1 (low expression of OXPHOS genes, OXPHOS-low), and group 2 (high expression of OXPHOS genes, OXPHOS-high), Log-Rank test. c Overall survival in the ESR1-altered tumors according to the 16 OXPHOS genes signature: group 1 (low expression of OXPHOS genes, OXPHOS-low), and group 2 (high expression of OXPHOS genes, OXPHOS-high), Log-Rank test.

Discussion

Although progress has been made in the management of patients with ER + / HER2-negative advanced BC, the prognosis remains poor when endocrine resistance to aromatase inhibitors occurs. It is therefore crucial to identify the molecular mechanisms underlying the selection of this treatment. In our analysis of patients with advanced ER + / HER2-negative BC progressing on aromatase inhibitors, we first identified ESR1 fusions or mutations as the most common alteration (49%). These alterations were not mutually exclusive with those of PIK3CA, MYC, NF1 or ERBB2. We then examined the transcriptomic profile of patients with an ESR1 alteration that occurred during treatment and identified differential expression of the ER-dependent reactive pathway. We generated a signature comprising 17 genes of this pathway, which was able to predict the presence of an ESR1 alteration with a sensitivity and specificity of 81.3 and 84.3%, respectively. Finally, we examined the expression profile of everolimus-sensitive tumours, as our patients were subsequently treated with everolimus in the SAFIRTOR trial. We identified the overexpression of oxidative phosphorylation pathway genes as significantly associated with everolimus resistance.

The first significant result of our work concerns the analysis of the landscape of molecular alterations that occur under the therapeutic pressure induced by aromatase inhibitors. In the study by Razavi et al. [7], ESR1 pathogenic variants were also the most frequent alterations. The emergence of an ESR1 variant is indeed an important mechanism of resistance to aromatase inhibitors, and its prevalence in other publications (20-40%) is slightly lower than that reported in our work [7, 36]. This difference could be explained by the fact that our cohort is homogenous in terms of previous hormone therapy, as all 65 patients included were pretreated with aromatase inhibitors only. Notably, ESR1 fusion genes were identified in two samples, one of which harbored two distinct fusions. This may represent an instance of convergent evolution, wherein two distinct clones independently acquire distinct yet functionally equivalent resistance mechanisms in response to selective therapeutic pressure [37]. The expression profile of ESR1 fusion tumours and ESR1 mutant tumours were similar, suggesting that the fusions are activating and responsible for endocrine resistance. A recent study has also shown that certain ESR1 fusions, located within the functional domain of the protein, are a recurrent mechanism of resistance to AIs [38]. ESR1 is a transcription factor of the nuclear receptor superfamily that mediates the effects of estrogen ligands and promotes growth and cell survival [39]. A study recently showed by performing global proteome analysis cell proliferation (e.g. mTORC1 signaling, MYC targets) and metabolism (e.g. oxidative phosphorylation, OXPHOS) were enriched in mutant models in estrogen deprivation conditions [40].

Additionally, we identified a 6% rate of ERBB2 mutations. ERBB2 activating mutations have been identified as a mechanism of acquired resistance to estrogen receptor targeted therapies in HR+ mBC due to cross-talk between the two signaling pathways [41, 42]. The actionability of these acquired mutation was recently confirmed by the SUMMIT trial, which demonstrated the efficacy of neratinib, a pan-HER kinase inhibitor, in combination with fulvestrant and trastuzumab in this context [43]. Another member of the mitogen-activated protein kinase (MAPK) signaling cascade is NF1, of which we reported a 6% alteration rate in our cohort. Alterations resulting in NF1 silencing have been described in the AI-resistant metastatic ER + /HER2- BC setting, leading to cyclin D1 overexpression and increased RB1 phosphorylation in preclinical models [44]. A current trial (phase II ComboMATCH NCT05554354) is comparing the use of fulvestrant with binimetinib plus fulvestrant in ER + /HER2- mBC in patients with inactivating or presumed inactivating NF1 alterations in their tumors. These studies raise the question of systematic tumor molecular profiling for endocrine resistance in routine practice. It should be noted that Baselga et al. reported a mutual exclusion between NF1 inactivating mutations and ERBB2 activating mutations with ESR1 alterations, which we did not find in our present cohort.

Transcriptomic profile of tumors harboring an alteration in ESR1 compared to WT tumors, showed several up-regulated genes, including Growth Regulation by Estrogen in Breast Cancer 1 (GREB1), which is a cofactor within the EZH2-ERα-GREB1 axis. Although the exact role of GREB1 is not fully understood, it has been implicated in endocrine resistance [45]. One possible explanation is its involvement in chromatin remodeling, leading to topographical interactions both within and between topologically associating domains (TADs), which are frequently altered in ER+ endocrine-resistant BC cells [46]. Together, these mechanisms contribute to altered expression of ER-regulated genes, consistent with the dynamic remodeling of ER pathways that accompanies the development of endocrine resistance. Most interestingly, we were able to identify a signature based on our RNA sequencing analysis with a sensitivity of 81.3% and a specificity of 84.3% for predicting the presence of an ESR1 alteration. To our knowledge, this is the first signature reported in the literature in this specific setting.

None of the patients in this study received a CDK4/6 inhibitor as a first-line treatment for metastatic disease, as CDK4/6 inhibitor approval in France occurred after the SAFIRTOR study. Therefore, these results may not be fully applicable to patients more recently treated with a combination of an AI and a CDK4/6 inhibitor. Additionally, while the BOLERO-2 study [19] demonstrated the benefit of exemestane (a steroidal AI) plus everolimus after progression on an AI, currently preferred options following AI progression, particularly in the presence of an ESR1 mutation, may include the combination of Selective Estrogen Receptor Degraders (SERDs) as fulvestrant or elacestrant (if available) with everolimus.

Finally, we observed an overexpression of oxidative phosphorylation (OXPHOS) pathway genes in everolimus-resistant tumors. Oxidative phosphorylation (OXPHOS) is the primary metabolic pathway through which cells generate ATP via the electron transport chain (ETC) in mitochondria [47], playing a crucial role in energy homeostasis and redox balance [48]. In cancer, metabolic plasticity is a key mechanism driving therapeutic resistance, with some tumors exhibiting a shift towards increased OXPHOS activity to support survival under therapeutic pressure (the “Warburg Effect”). A recent study using genomic analyses of patient tumors, metastatic patient-derived xenografts (PDX) and isogenic cell lines showed that some metastatic ER + BC were highly dependent on OXPHOS [49]. Other studies have suggested that metabolic dependencies, including increased mitochondrial respiration, can drive resistance to PI3K/AKT/mTOR-targeted therapies, highlighting a potential avenue for therapeutic intervention [50]. The identification of this pathway as a mechanism of treatment escape is exciting as it opens the way to new potential treatments, the efficacy of which will need to be evaluated in clinical trials. Indeed, in the aforementioned study, IACS-010759, an inhibitor of complex I of the electron respiratory chain, has shown an increased ability to halt the growth of endocrine and palbociclib-resistant PDX. In addition, a number of Phase 1 studies have evaluated IACS-010759 in humans [51]. The development of this drug is still limited by its inherent neurotoxicity. Further trials are therefore needed to improve its development in solid and hematological tumors. Last but not least, we identified a 16-gene signature that can predict response to everolimus. It is to note that the vast majority of these 16 OXPHOS genes are part of subunits II, IV and V of the mitochondrial respiratory chain [52].

This study has several limitations, including the limited number of patients (n = 65) and the lack of a validation set to test our two gene signatures of clinical interest. Furthermore, these results may not fully apply to patients more recently treated with a combination of AI and CDK 4/6 inhibitors. None of our patients neither received PI3K inhibitors. These results warrant further investigation in other more recent translational studies and larger prospective clinical trials.

Conclusions

Molecular analysis of tumour samples from the SAFIRTOR trial indicates that endocrine resistance in advanced or metastatic HR + BC is largely driven by ESR1 alterations mediated by the ESR-mediated pathway. In addition, these tumors acquire an OXPHOS phenotype that confers resistance to everolimus, and may represent a specific subgroup for which new targeted therapy is warranted.

Author contributions

Please use the following template for author contributions: Solenn Barraud wrote the manuscript; Sophie Vacher, Fabrice André, Thomas Bachelot, François Legrand, Marta Jimenez and Ivan Bièche designed the research; Camille Schiffler, Jennifer Wong, Laura Courtois, Valery Attignon, Isabelle Soubeyran, Etienne Rouleau, Mario Campone and Florence Coussy performed the research; Solenn Barraud, Sophie Vacher, Abderaouf Hamza, Nicolas Servant, Sylvain Baulande and Ivan Bièche analyzed the data.

Funding

SAFIRTOR trial was funded by NOVARTIS and UNICANCER.

Data availability

The datasets generated during the current study are available in the SRA database under accession, code SUB15219028.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41416-025-03083-5.