Evaluation of drug-specific multigene markers to predict aromatase inhibitor efficacy in advanced breast cancer: a cohort study from Danish Breast Cancer Cooperative Group

Background Even with patients fail to benefit from endocrine therapy (ET). A method that previously predicted other drugs various cancers was evaluated. Here multigene markers based on aromatase inhibitor (AI) effect in vitro were used for prediction of AI benefit in ER+ ABC patients. Simultaneously effects of long-term ET on predictive efficacy was evaluated. Methods The Drug Response Predictors (DRPs) are based on correlations between baseline gene expression and growth inhibition patterns of exemestane, anastrozole and letrozole, respectively, in the National Cancer Institute 60 cell lines. The genes were controlled for expression in 3,500 tumours. In a Danish Breast Cancer Cooperative Group cohort of 695 ABC patients with complete gene expression and time-to-progression (TTP) data, 414 received an AI as monotherapy. Hereof, 57 received anastrozole, 166 received exemestane, and 327 received letrozole. mRNA was isolated from archival formalin-fixed paraffin embedded tumour tissue and run on microarray and 60% of the tumours were from time of primary diagnosis. Medical records of the patients were assessed for TTP for all treatments given for ABC. Results The DRPs were tested in subsets 1) with no adjuvant ET and 2) with adjuvant ET. In 1) the anastrozole DRP predicted benefit of anastrozole (hazard ratio (HR) was 0.21 upper 95%-confidence interval limit (CI) 0.76, p=0.023) but not in 2). Dichotomised by a DRP of 50, the anastrozole DRP did predict benefit (HR=0.16, upper 95%-CI 0.75, p=0.026). Only in 1) the exemestane DRP predicted benefit of exemestane (HR=0.57, upper 95%-CI 1.00, p=0.0497). The letrozole DRP had no predictive value. Additionally, we tested each DRPs ability to predict other AIs. Only the anastrozole DRP predicted benefit of overall AI treatment, in 1) with an HR of 0.76 (upper 95%-CI 0.99, p=0.044) and in 2) with an HR of 0.71 (upper 95%-CI 0.92, p=0.015). The anastrozole DRP did though not predict benefit of letrozole. All tests are one-sided, alpha=5%. Conclusions Among the DRPs for AIs, the anastrozole DRP was strongest with clinically relevant prediction of TTP in AI treated ER+ ABC patients. Trial registration: ClinicalTrials.gov NCT01861496.

(ER+) disease and endocrine manipulation is a mainstay in their treatment. The principal source of oestrogens in postmenopausal women is aromatization of adrenal androstenedione and testosterone to oestrogens. Consequently, for postmenopausal women with ER+ breast cancer, treatment will often include an aromatase inhibitor (AI) (1). This removes a main source of growth stimulus for the oestrogen-receptor positive breast cancer. Three AIs are currently used: Anastrozole, letrozole, and exemestane. All are 3 rd generation AIs and while anastrozole and letrozole are reversible nonsteroidal inhibitors, exemestane is an irreversible steroidal AI (2). Response rates to AIs as monotherapy vary from 21-32 % for ER+ ABC patients (3)(4)(5). Recent successes of combining an AI or fulvestrant with a cyclin-dependent kinase (CDK) 4/6 inhibitor in ER+, human epidermal growth factor receptor 2 negative (HER2-) ABC has moved this combination to first and second line treatment (6,7). Even as the additional benefit of CDK4/6 inhibition to AIs is very impressive, biomarkers to identify primary endocrine resistance are important in the 20-25 % of patients experiencing early progression on these treatment regimens (6,8).

Predictive biomarkers
In ABC, ER and HER2 analyses have improved the ability to predict which patients to select for treatment with endocrine therapy (ET). Further improvement is needed as no biomarkers predictive of specific treatment outcome have been approved neither in early nor advanced stages of breast cancer, thus, it is currently not possible to select between fulvestrant, tamoxifen and the three AIs in clinical use. Research in the field includes measures of ER, ESR1, the coding gene for ER-α, HER2, phosphatidylinositol-4,5-bisphosphate 3-kinase catalytic subunit alpha (PIK3CA), the coding gene for interleukin 6 signal transducer (IL6ST) and multigene biomarkers. Specific mutations in ESR1 and PIK3CA are clearly linked to resistance to endocrine treatments (9,10). In a meta-analysis of largescale trials ATAC, BIG 1-98 and TEAM HER2-patients appeared to gain significantly greater benefit from treatment with adjuvant AIs compared to adjuvant tamoxifen while there was no difference between treatments in the HER2+ subgroup (11). Prediction of neoadjuvant letrozole response was obtained with 4 genes including IL6ST (12). Lastly, Ueno et al found that low levels of the multigene marker Oncotype Dx® recurrence score (RS) predicted benefit of exemestane in a neoadjuvant trial (13).

In vitro based markers of aromatase inhibition
For postmenopausal patients a large part of the oestrogen production is believed to arise in the tumour itself (14,15); and aromatase expression is confirmed in both stromal cells and carcinoma cells of breast cancer (16). Immunohistochemistry staining for aromatase in primary tumours did, however, not predict clinical outcome of treatment with AIs in the advanced setting (17). Similarly, aromatase expression in ABC patients did not predict efficacy of letrozole (18).
In a study by Sonne-Hansen and Lykkesfeldt oestrogen-dependent MCF-7-breast carcinoma cells survived on serum with testosterone, the substrate for aromatase to make oestrogen. They also showed how AIs exhibit a cytotoxic effect in these cultured breast cancer cells (19). Another recent study showed intra-tumoral amplification of aromatase following clinical resistance to aromatase inhibition (20). Despite a distinct mode of action and well described in vitro systems a predictive biomarker for AI efficacy other than ER has not been developed.

New AI predictors
In early breast cancer, the use of complex multigene biomarkers for prognostic purposes has been widely endorsed by the medical oncology societies for clinical practice (21)(22)(23)(24)(25)(26)(27). Similar to the complex prognostic markers based on multiple genes we developed multigene profiles specific to AI treatment: Drug Response Predictors (AI-DRPs). These are based on data from the National Cancer Institute panel of 59 human cancer cell lines (NCI-60). In NCI-60 each tested drug shows a characteristic pattern of sensitive and resistant cell lines (28). Baseline gene expression and protein levels are measured on all cell lines (29).
AIs have been shown to exhibit cytotoxic effects in various cancer cell lines (30,31) and NCI-60 ( Fig.   1) (28). While NCI-60 has few breast cancer cell lines Knudsen et al previously showed that it was possible to predict clinical fulvestrant efficacy by use of a fulvestrant specific DRP from NCI-60. Here no additional benefit was obtained by testing in a specific breast cancer cell line panel (32). This may be explained by the presence of oestrogen and derivates in all 60 cell lines with surprising abundancy in melanoma and leukaemia cell lines (33).
We thus hypothesize that AI sensitive cancer cells in NCI-60 has a characteristic pattern of gene expression that differs from AI resistant cancer cells and that this pattern, in a modified form, can predict benefit of AIs in clinical practice. By following REMARK guidelines (34) we assess if there is an association between predicted sensitivity to AIs exemestane, anastrozole or letrozole and time-toprogression (TTP) in a Danish Breast Cancer Cooperative Group (DBCG) cohort of ABC patients.

Development of the DRP
The method of developing drug response predictor profiles has been described before in development of a DRP for fulvestrant (32). At the National Cancer Institute cell lines are subjected to a drug and growth inhibition values (GI50) are measured and published online (28). These cell lines are referred to as the NCI-60. 49 of these were subjected to anastrozole, 66 to exemestane and 57 to letrozole, searched on NCI-60 database with NSC numbers 719344 (anastrozole), 713563 (exemestane), and 719345 (letrozole) (35). Included are 6 cell lines derived from breast cancer (only 5 in the anastrozole and letrozole treated subsets). On the same cell lines, with no drug present, baseline gene expression has been measured on microarray (29). We used correlations between baseline gene expression and patterns of growth inhibition of anastrozole, exemestane, and letrozole ( Fig. 1). Anastrozole, exemestane, and letrozole were tested individually and unique patterns of sensitivity were seen for each drug. When letrozole was applied to the cell lines only 7 out of 57 cell lines had differential GI50 values. The letrozole predictor was thus not expected to perform as well as the other predictors.
Genes that in the cell lines were correlated to a low GI50 of anastrozole were considered sensitive to anastrozole and were therefore biomarkers of sensitivity and retained in the anastrozole DRP. Genes correlated to high GI50 levels were considered resistant to anastrozole and retained as biomarkers of resistance. This was repeated for exemestane and letrozole to produce AI DRPs specific for each AI.
To reduce noise caused by false positive genes, a filter was applied, securing that the selected genes were biologically relevant as the filter removes genes only active in cell lines. The filter consisted of 3,500 tumours of mixed origin (Fig. 1).
For each patient, three DRP scores were calculated as the difference between the mean of all sensitive genes and the mean of all resistant genes for each of the AIs tested. This was compared to a reference population of 819 breast cancer patients. The DRP scores were expressed as a percentile of the reference population. Higher scores predict higher likelihood of sensitivity to anastrozole, exemestane or letrozole, respectively. These final AI drug response predictor scores were tested blinded to outcome. This system has been assessed and validated in various cancer types with several anticancer agents (32,36,37).

Patients
Totally, 1,199 patients with locally advanced or metastatic breast cancer have given informed consent to be enrolled in a screening cohort selecting patients to a phase II trial with liposomal cisplatin (LiPlaCis), the DBCG screening cohort. The phase II trial with LiPlaCis is registered with ClinicalTrials.gov as NCT01861496 (38). Data from medical records for patients enrolled in the DBCG screening cohort was used to test markers for prior treatments as previously described by Buhl and colleagues (39). In brief, patients at least 18 years of age with histologically confirmed locally advanced or metastatic adenocarcinoma of the breast and performance status < 2 with no other malignancy within 5 years prior of the enrolment were eligible. Clinical data was extracted retrospectively from medical records of each patient and included among more information on duration of each treatment for ABC and information on reason for treatment change defined as either progressive disease, toxicity, physician's decision, patient's decision or long-lasting stable disease.
The study was approved by the Regional Committee on Health Research Ethics for the Capital Region Denmark (ID HGH-2016-097). Informed consent for this part was not obtained as the research did not have implications on the health of the research subjects as accepted by the Regional Committee on Health Research Ethics for Capital Region Denmark.

Specimen characteristics
Archival formalin-fixed paraffin-embedded (FFPE) tissue from the patients' breast cancer was used. Of 414 unique patients treated with anastrozole, exemestane, or letrozole tissue blocks were from the original surgery (n = 247), any surgery performed after 4 months from diagnosis including biopsies from metastasis (n = 76) or unknown (n = 91). Sampling was done from tissue block (cores or slices) or from biopsy (core from needle). Tissue was retrieved from the newest possible tumour specimen with control of tumour cell content, necrosis and haemorrhage by certified pathologist Eva Balslev.
Only blocks with at least 5 % tumour cells were used.

Assay methods
In the sample from each patient total RNA was extracted using RecoverAll TM Total Nucleic Acid Isolation Kit for FFPE samples (Ambion, Austin, Texas) according to manufacturer's protocol. Total RNA was then amplified and applied to microarray Genechip® Human Genome U133 Plus 2.0 Array (Thermo Fisher Scientific, Waltham, MA, USA). Samples with in vitro transcription yield below 500 ng/ µL were discarded.

Study design and statistical analysis plan
Study design was retrospective with no matching control. Prior to unblinding a statistical analysis plan was specified. The plan was to assess any associations of the DRP profiles for anastrozole, exemestane, and letrozole to clinical outcomes on treatment with the same drug. Primary endpoint was time to progression (TTP) defined as time from start of treatment to progression. When primary endpoint was not reached patients were censored on date of latest contact. Other censoring events were death or other outcome such as toxicity, physician's decision or patient's decision to terminate treatment.
Variables considered in the multivariate analysis were DRP, age, adjuvant endocrine treatment, HER2 status, ER status, performance status and number of metastatic sites.
The assessment was done with the cox proportional hazards model with stratification for treatment line and adjusted for previous adjuvant endocrine therapy including an interaction term between this and the DRP. Prespecified in the statistical analysis plan was scoring the DRP as continuous variable with hazard ratio for a difference of 50% points. We also evaluated the DRP grouped by below or above 50%. Only complete cases were used.
As the hypothesis was that high levels of the DRP predict benefit of a drug, all confidence intervals are one sided (upper limit) with a 5% significance level. Model assessments of proportionality and linearity were done using martingale residuals.
Post-hoc analysis included testing each DRP and DRPs for fulvestrant and tamoxifen on all AI treated patients as a random effect for each AI treatment. This was done to assess the specificity and to identify the most robust predictor for future purposes. Other post hoc analysis included predictive models based on the cox proportional hazards model for both the anastrozole DRP in all AI treated and in patients treated with only anastrozole or exemestane.

Baseline patient characteristics
Patients included in the study population were diagnosed with primary breast cancer between 1979 and 2015 and received an AI in the locally advanced or metastatic setting between June 1997 and November 2016.
In total, 695 ABC patients were evaluable for time-to-progression and had enough tissue for gene expression (Fig. 2). Hereof 414 had received an AI at any given time-point for metastatic disease of which 57 received anastrozole as monotherapy, 166 received exemestane as monotherapy, and 327 received letrozole as monotherapy. This included 19 that received both anastrozole and exemestane, 74 that received letrozole and exemestane, 11 that received letrozole and anastrozole and 16 that received letrozole, exemestane, and anastrozole as monotherapy at separate time points. Anastrozole was administered as the 2 nd treatment line (median) for advanced disease and median TTP was 29.3 months, exemestane was administered as 3 rd treatment line (median) for advanced disease and median TTP was 8.5 months, whereas letrozole was administered as 1 st treatment line (median) and median TTP was 22.2 months (Table 1).
Among all patients treated with anastrozole 32 patients of 56 had received adjuvant ET hereof 50% had received tamoxifen alone (

Predictive value of the markers
The patients assessed were those treated with an AI in any treatment line at any time for ABC. We grouped the patients into those that had received prior adjuvant ET or not, stratified for treatment line and with random effect for treatment (anastrozole, exemestane or letrozole) when assessing the DRP in all AI treated patients. The following variables were adjusted for in the multivariate model: adjuvant ET, treatment line, ER status, number of metastatic sites (1, 2, 3, >3), age (<50, >50-60, >60-70, >70) and performance status (0-1 vs 2-). Each of the AI DRPs were tested on the treatment they were developed on, anastrozole DRP was then as a primary objective tested in the anastrozole treated patients, the exemestane DRP in the exemestane treated patients and so forth. Secondarily we evaluated each of the DRPs in the entire subset of AI treated patients from the DBCG cohort (Table   3).

The anastrozole DRP
In the anastrozole treated patients that had not received adjuvant ET the anastrozole DRP predicted TTP with a HR of 0.21 (upper 95%-CI 0.76, p=0.023) ( Table 2). The anastrozole DRP was scored continuously, with HR presented for a 50-point difference of the DRP score. The performance of the anastrozole DRP remained significant in a multivariate model. When dichotomising the anastrozole DRP by 50 the HR was 0.16 (upper 95%-CI 0.75, p=0.026) (Fig. 3) in the patients that had not received adjuvant ET. In the subset of anastrozole treated patients that had received adjuvant ET previously the anastrozole predictor did not predict TTP (HR 1.05, p=0.54) ( Table 2 and Additional File 1: Fig. S1).

The exemestane DRP
In the exemestane treated patients naïve to adjuvant ET, the exemestane DRP predicted TTP with an HR=0.56 (upper 95%-CI 1, p=0.0497) ( Table 2). This was robust through all treatment lines with exemestane given for advanced disease and in a multivariate model with an HR of 0.47 (95%-CI upper-limit 0.87, p=0.023). Otherwise only performance status was statistically significant. When patients in this subset were dichotomised by exemestane DRP > 50 versus exemestane DRP < 50 HR was 0.43 (upper 95%-CI .79, p=0.022) in a cox proportional hazards model with stratification for treatment line (Fig. 4). No predictive effect of the exemestane DRP was found in patients who had received adjuvant ET (Table 2 and Additional File 1: Fig. S2). Some early results for the exemestane DRP was presented at American Society of Clinical Oncology Annual Meeting 2017 (43).

The letrozole DRP
No effect of the letrozole DRP was present in the letrozole treated patients ( Table 2).

The DRPs in all AI treated patients
To assess any ability to predict treatment outcomes of treatment with other AIs we did additional post-hoc analysis evaluating each predictor in each subset of AI treated patients.
The anastrozole DRP predicted benefit of treatment with either anastrozole, exemestane or letrozole with an HR of 0.76 (upper 95%-CI 0.99, p=0.044) in the subset that had not received adjuvant ET (  Fig. 3, and in Additional File 1: Fig. S3a and S3c).
The anastrozole DRP also predicted TTP in the subset of AI treated patients that had received previous adjuvant ET. Scored continuously HR for a 50-point difference was 0.71 (upper 95%-CI 0.92, p=0.015) ( Table 3 and Additional File 1: Fig. 1 and S3b and S3d). However, the anastrozole DRP was not predictive of treatment with letrozole in either subset ( Table 3).
The exemestane DRP did not predict benefit of treatment with neither anastrozole nor letrozole (pvalues ranged from 0.26-0.98) ( Table 3). The letrozole DRP appeared to predict some benefit of exemestane in the subset that had received adjuvant ET previously (HR=0.51, p=0.004) but was otherwise not predictive of TTP (Table 3).
As the anastrozole DRP had an improved performance in both the anastrozole treated and the exemestane treated subsets and a better overall performance than the exemestane and letrozole DRPs, the anastrozole DRP was chosen for further evaluation.

Predictive models
Based on the multivariate cox model, a predictive model was specified for all AI treated in the subset with no previous adjuvant ET. The model predicted outcomes for a given DRP level of the anastrozole DRP, no other variables were included in this predictive model. In all AI treated estimation of median TTP for a DRP of 10 was 16 months (95%-CI 13-20 months), for a DRP of 50 the estimation was 19 months (95%-CI 16-22 months), and for a DRP of 90 estimation was 23 months (95%-CI 18-30 months). As there was no effect of the anastrozole DRP in the letrozole treated patients we specified a model post-hoc without the letrozole treated patients as well. In this model the estimated TTP for an anastrozole DRP score of 10 was 6.2 months (95%-CI 3.5-17.8 months), for a DRP of 50 the estimation was 17.8 months (95%-CI 8.5-25.3 months) and the estimated TTP for a DRP of 90 was 27.0 months (95%-CI 18.0 months -not possible to estimate).

Comparison of predictors
As an explorative analysis we compared the prediction of the AI DRPs to DRPs based on fulvestrant and tamoxifen on the AI treated patients (32). Neither the fulvestrant DRP nor the tamoxifen DRP predicted any overall benefit of AI treatment (p-values ranged 0.35-0.93). Subgroup analysis showed that the fulvestrant DRP was able to predict long TTP in the letrozole treated ABC patients that had received adjuvant ET (for continuous scoring with a difference of 50 HR was 0.42, p=0.0014).
The Pearson correlation between the anastrozole DRP scores and the exemestane DRP scores was modest, r=0.25. Thus, the predictors appeared to identify different patients as benefitting treatment which was also obvious in the differential effect of the two DRPs.

Discussion
Postmenopausal patients with ER+ breast cancer who have not received an AI in the adjuvant setting are generally recommended an AI following distant metastases. Our study suggests that those who are likely to benefit from exemestane or anastrozole may be identified by DRPs based on exemestane or anastrozole, respectively. The basis for the DRP scores is gene expression using FFPE tumour tissue from either the primary tumour (60%) or unknown or a later recurrence (40%) (no further details were available). Either may have been excised from the patients many years prior to initiation of the treatment we are assessing. Prediction across all AI treatments was attempted and the anastrozole DRP appeared to predict benefit of exemestane in addition to anastrozole in both the subset treated with adjuvant ET and in the subset that did not receive adjuvant ET. The predictive effect of the anastrozole in exemestane treated patients points to a similar mechanism of action in the AIs.
However, no prediction of benefit of letrozole was possible with either of the AI DRPs. We did not find any interactions between letrozole treatment and baseline characteristics nor any issues with proportionality that could explain this.
The assessed patients are treated with anastrozole, exemestane or letrozole in any treatment line at any time for advanced disease and the effects of the DRPs were not reduced with later treatment lines or with other endocrine treatments for advanced disease.
We planned and conducted the analysis stratified by prior adjuvant ET status because we had hypothesised that long term exposure to endocrine therapies would alter gene expression in a way that we could no longer use samples from before the exposure. The tissue blocks used in the study were obtained from archival samples of the current breast cancer from each patient. The samples represent a pre-treatment baseline gene expression level and patients received 3 years (median) of adjuvant ET. The inability of the exemestane DRP to predict outcome when patients had received adjuvant antihormonal therapy could be caused by the anticipated substantial change in genes in the tumour after several years of endocrine treatment (44). To assess whether the exemestane DRP is useful in patients with previous adjuvant endocrine treatment using a new biopsy after adjuvant treatment will be necessary.
Letrozole had a marginal effect on the cell lines in NCI-60 as only 7 of the 57 cell lines had differential GI50 values. Therefore, it was not very likely to succeed as a predictor of benefit in patients. We conclude that new experiments at NCI-60 or on other cell lines are needed to improve this predictor.
Additional examinations and validations in the letrozole treated patients are necessary as our current results are conflicting: the fulvestrant DRP appeared to predict in letrozole treated patients that had received adjuvant ET.
Limitations of the current study include the retrospective and non-randomised study design and thus lack of an untreated control group. A comparator group could enlighten whether the markers are predictive of AI treatment or prognostic in this subset of patients. Another limitation is the small sample sizes in subgroups but also in the anastrozole treated subset altogether. Strengths include that this is a blinded, independent validation of locked-down versions of the DRPs.
As approximately 20% of ER+ patients are de-novo resistant to AI treatment accuracy in patient selection can be improved (8). In our cohort of AI treated ABC patients all but 9 patients are ER+. As the DRP from anastrozole can discriminate patients with differential benefit of anastrozole and exemestane in this cohort, it is evident that ER status is not enough to select patients for treatment.
However, no biomarkers predicting choice of specific endocrine treatment are approved for clinical use.
Currently the treatment algorithm for advanced hormone-receptor positive disease patients is rapidly changing with CDK4/6 inhibitors and fulvestrant or AI being current first line (6). Fulvestrant appears superior to AIs (45,46), but we hypothesize that some patients may gain more effect by AIs than fulvestrant, as the marker predicts differential outcome between the two drugs in an all ER+ population (Fig. 5). A biomarker to predict response to either fulvestrant or AIs would be of great clinical use. A similar DRP for fulvestrant has been shown to predict benefit of fulvestrant in the DBCG cohort and retrospectively in a neoadjuvant trial with fulvestrant (32,47). As the DRP system has previously been found to be predictive in settings with multiple drugs (48,49), this system could then potentially point to which antihormonal treatment would be preferable: fulvestrant or AIs in the first line setting of combinations with CDK4/6 inhibitors.

Conclusion
The anastrozole DRP may predict benefit of exemestane and anastrozole in ABC patients. The prediction was best in patients that had not received adjuvant ET. No prediction of letrozole benefit was found.
In this cohort predictions of benefit were made from archival FFPE tumour tissue most of which were

Consent for publication
Not applicable.

Availability of data and material
The datasets analysed during this study are available through DBCG on reasonable request (dbcg@dbcg.dk).

Competing interests
The authors declare the following potential competing interest: authors IKB, ASKB, UHB, SK, and PBJ declare past or present primary employment or leadership position in a company (Oncology Venture) that has a potential to benefit from these results. Authors IKB (immediate family), ASKB, IJC, UHB, SK, and PBJ declare ownership in Oncology Venture. Author IJC declares advisory role to Oncology Venture. Author BE declares funding to institution by Oncology Venture and other outside the submitted work. Authors ASK and EHJ declare consultancy roles outside the submitted work. Authors

Supplementary Files
This is a list of supplementary files associated with this preprint. Click to download.