Introduction
Aromatase inhibitors (AIs), such as anastrozole or letrozole, block the synthesis of estrogen [
1]. AIs are the standard of care for the treatment of estrogen receptor (ER)-positive breast cancer in postmenopausal women [
2]. Estrogen deprivation has a rapid effect on transcriptional profiles, with substantial gene expression changes identified after 15 days of treatment [
3,
4]. The most frequently upregulated pathways are those associated with focal adhesion, actin cytoskeleton and inflammation, while the most frequently downregulated pathways are those related to proliferation, growth and ER transcription [
5].
Acquired or
de novo resistance to AIs is common [
6], and multiple putative mechanisms of resistance to AI therapy have been proposed. These include intrinsic resistance of tumors to estrogen, aromatase-independent estrogenic hormones, signal transduction by non-endocrine pathways and selection of hormone-insensitive clones during AI therapy (reviewed by Miller
et al. [
7]). A number of potential biomarkers of resistance have been suggested, including overexpression of human epidermal growth factor receptor-2 (HER2), Cyclin E1, hypoxia-inducible factor (HIF)1α and p44/42 mitogen-activated protein kinase (MAPK) [
8]. These biomarkers, however, still require validation in independent cohorts [
7] or are unlikely to account for resistance to AIs in the majority of tumors [
9]. The identification of robust predictive biomarkers for resistance or sensitivity to AIs is therefore a research priority.
The observed changes in transcription following treatment with AIs led to the identification of gene expression signatures in pre-treatment tumor samples reported to be predictive of response to AIs, as measured by a decrease in tumor volume [
6,
10]. To our knowledge, neither of these signatures has been validated in a larger independent cohort. The challenges of translating predictive gene expression signatures into clinically useful tools are now well-recognized [
11]. These include, but are not limited to, the facts that 1) resistance to a given agent may be mediated through multiple distinct pathways in different tumors, 2) the low sensitivity of microarray platforms for low-level changes in expression or for changes in non-modal clones may not detect the mechanism, and 3) resistance to an agent may not manifest in transcriptomic changes, but may be mediated through mutations or epigenetic aberrations that do not result in overt transcriptomic changes.
Gene amplification is a common mechanism of oncogene activation in cancer [
12]. There are multiple reports describing the association between specific gene amplifications and resistance to various anti-cancer therapies. For example, in breast cancer, resistance to tamoxifen is associated with
FGFR1 amplification [
13], while amplification of
CCNE1 [
14] and
IGF-1R [
15] are associated with resistance to trastuzumab. Further examples abound in other tumor types, such as the association of
ERBB2 [
16] and
CRKL [
17] amplification with resistance to anti-epidermal growth factor receptor (EGFR) targeted agents in non small-cell lung cancer and
YAP amplification with resistance to doxorubicin in hepatocellular carcinoma [
18].
Alternative approaches to identifying biomarkers of resistance to therapy include the use of genome-wide copy number profiling microarrays to compare the patterns of copy number aberrations (CNAs) between responders and non-responders. This approach has identified genomic loci associated with response to various chemotherapeutic agents in ovarian carcinoma [
19], large B-cell lymphoma [
20] and colorectal carcinoma [
21], to name but a few. Amplified regions frequently encompass multiple genes and not all genes within an amplicon are overexpressed and of functional significance [
22].
By integrating genome-wide copy number profiling data and gene expression data, lists of genes associated with response to specific therapies can be enriched for biologically relevant targets (for example, the identification of
FGFR1 amplification as a modulator of tamoxifen response [
13]). More recently, publication of the Cancer Cell Line Encyclopedia [
23] and the Genomics of drug sensitivity [
24] datasets has demonstrated the power of integrative genomic and functional genomic approaches in identifying determinants of response to targeted therapies. To date, there are limited genome-wide data identifying CNAs that are associated with response to AI therapy measured by Ki67 as an intermediate endpoint. Ellis
et al. [
25] compared whole-genome analysis in resistant versus sensitive tumors, using Ki67 at surgery as the index of response; however, the study focused on mutational background and somatic structural variations and not specific copy number and expression changes.
The aims of this study were to 1) relate patterns of copy number aberrations to molecular and proliferative response to endocrine treatment, 2) study differences in the patterns of copy number aberrations between breast cancer samples pre- and post-AI neoadjuvant therapy, and 3) identify putative biomarkers for resistance to neoadjuvant AI therapy using an integrative analysis approach.
Discussion
In this study, an integrative analysis of copy number profiling, gene expression microarray and Ki67-based AI response data was performed using data from samples in two cohorts of neoadjuvant AI therapy in postmenopausal patients with ER-positive early breast cancer. No significant differences in the frequency of gene copy-number aberrations were found when pre- and post-AI therapy samples were compared. This is to be expected given that there is evidence for numerous mechanisms of resistance to AIs many of which would not be expected to be dependent on gene copy-number changes [
7]. Non-recurrent differences in copy number at specific loci (that is, 1p11.23-p11.22, 1q31.1-q41, 3q13.11, 8p11.23-p11.22 and 12q12) between samples before and after 3 months of letrozole therapy were observed in 6 of 19 studied cases (Figure
2B). This observation suggests that the selective pressure applied by AI therapy may result in the selection of non-modal clones, causing enrichment or loss of cells harboring specific amplicons, and that resistance to AIs may constitute a convergent phenotype [
11]. An alternative explanation, however, is that these differences in gene copy-number profiles may be merely a manifestation of spatial intra-tumor genetic heterogeneity. Indeed, this study highlights the need to interrogate intra-tumor genetic heterogeneity for a full understanding of the mechanisms of resistance to specific agents, using combinations that can overcome the challenges that may be posed by intra-tumor spatial genetic heterogeneity. All the same it should be noted that in this study few differences were noted in the copy number landscape between pre-treatment and on-treatment cores: any confounding degree of genetic heterogeneity that is dependent on the small amount of a tumor sampled in cores would have been revealed by that comparison.
Overlaying gene expression data with copy-number profiling data identified a set of 628 genes, which are significantly overexpressed when amplified, including genes at two commonly amplified regions (11q13.2-q13.4 and 17q12-q21.2). Nine of these 628 genes were also negatively correlated with the decrease in Ki67 expression after 2 weeks of AI therapy (a surrogate for response to AIs). These clinical data are reflective of pre-existing,
de novo mechanisms of resistance. To parallel these observations we assessed the possible importance of the genes in a panel of ER+ cell lines, two of which harbored the amplification and two of which did not. We prioritized the functional analysis of three of the nine genes based on differential expression between MCF7 and the LTED derivative (a model for acquired AI resistance), followed by functional validation. This pipeline identified
CHKA as a gene that is significantly overexpressed when amplified, and when amplified is associated with a poor Ki67 response to AI therapy. Mechanistic investigations revealed that CHKA expression modulates ER transcriptional activity via AKT and S6 phosphorylation, but independently of p90RSK activity, which resulted in a reduction in cell-cycle progression markers. Antagonizing ER signaling has been shown to attenuate cyclin-dependent kinase (CDK)/cyclin complexes at multiple levels [
46]. Furthermore, ER modulates transcription of cyclin D1. Hence suppression of ER signaling leads to inhibition of CDK activity and the maintenance of Rb in a phosphorylated and active state inhibiting progression to S-phase.
CHKA is located at 11q13.2 and encodes the protein choline kinase alpha (CHKA), which catalyses the phosphorylation of choline as the first step of the Kennedy (phospholipid synthesis) pathway [
47]. Choline phosphorylation by CHKA has been shown to be upregulated in many cancer types, including breast, lung, colorectal and prostate cancer [
48]. In our study we identified
CHKA amplification in 4% of cases, confirming previous reports [
49,
50]. Amplification at 11q13-q14 is a complex event. It is currently accepted that at least four cores of independent amplification are found at this locus [
51].
CHKA lies between the smallest regions of amplification of the first and second core of the 11q13-q14 amplicon. Curtis
et al. identified a high-risk group of ER-positive luminal tumors with amplification of 11q13/14 (Int2). This group harbors amplification of several genes such as
CCND1 and
EMSY and is in close proximity to the region containing CHKA [
49]. In the large METABRIC study of CNV and gene expression CHKA amplification does not associate with survival but its overexpression did correlate with poorer survival. It should be noted, however, that analyses of survival are generally unhelpful for detecting the impact of response or resistance to a particular treatment.
Using RNAi-mediated silencing and ERE-reporter techniques, the role of CHKA on ER-driven proliferation was characterized in this study, highlighting the importance of functional characterization of genomic and transcriptomic aberration using appropriate phenotypes as experimental readouts in multiple models. In this case, for example, previous studies demonstrated a significant effect of CHKA silencing on proliferation in MCF7 cells [
52]. However, those experiments were performed in complete growth medium; in the same study, silencing CHKA in serum-starved MCF7 cells produced no difference in proliferation, and an interaction between CHKA, EGFR and c-Src was demonstrated and found to be required for the pro-proliferation effect of CHKA. Assessment of the MCF7-LTED cell line showed that whilst they do not harbor amplification of CHKA the transcript level is significantly increased. Furthermore, this cell line shows elevated levels of both the EGF canonical pathway and c-Src [
53]. These data provide further support for the proposed mechanism of CHKA on ER-driven proliferation proposed in this study, which warrants further investigation.
This study has a number of limitations. First, it was limited by relatively small sample size. Although the two trials utilized in the study recruited over 160 suitable patients, only a subset of these samples were available in sufficient quantity for use in this study. The use of the neoadjuvant setting with Ki67 change as its primary endpoint provides substantially greater statistical power than a similar sized study of adjuvant therapy: with the former there is a direct readout of a patient’s response or resistance to treatment, which is not the case with the latter where patients are free of clinically detectable disease by which to judge response. It is important to note that Ki67 has been validated as an intermediate marker of long-term benefit from endocrine treatment [
27] and, in this respect it is a better endpoint than clinical response per se. However, differences in methodology for Ki67 immunohistochemistry and scoring between the two trials precluded combination of Ki67 response data. Likewise, given that different platforms were used for gene expression profiling in the two trials, this study was limited to the largest dataset (FAIMoS trial, n = 47) for the integrative copy number and gene expression analysis.
It should be noted that this study aimed to identify genes that have pathoclinical significance when both amplified and overexpressed. It is notable that there are reverse associations between cyclin-D1 expression with prognosis and resistance to tamoxifen or anastrozole according to whether
CCND1 is amplified or not (overexpression and amplified, poor prognosis; overexpression and non-amplified, good prognosis [
54]). Thus the approach taken in this study may not identify genes that associate with clinical phenotype only according to their degree of expression.
Competing interests
The authors declare that they have no competing interests.
Authors’ contributions
ELK and PW performed functional experiments, analysed the data and drafted the manuscript. HA extracted samples and performed aCGH experiments. AM and ZG performed the analysis of the aCGH data and statistical analysis. RR performed functional experiments. AR performed functional experiments. PO and AN performed the histophatology analysis. LR, AL, WRM and JMD provided the Edinburgh patients and extracted DNA. JSR and AKD were involved in conception and design. LAM was involved in conception and design, functional experiments and drafting of the manuscript, and MD conceived the study, and was involved in design and drafting the manuscript. All authors read and approved the final manuscript.