A Prognostic Model Based on PAM50 and Clinical Variables (PAM50MET) for Metastatic Hormone Receptor–positive HER2-negative Breast Cancer

Predicting long-term good prognosis in hormone receptor–positive (HR⁺)/human epidermal growth factor receptor 2 (HER2)-negative early breast cancer is of clinical value for de-escalating chemotherapy. To accomplish this, both clinical parameters and gene expression–based data are needed (1). First-generation and validated prognostic tests, such as OncotypeDX (2, 3) and MammaPrint (4), focus on gene expression data only in selected groups of patients. Second-generation tests (5, 6), however, integrate clinicopathologic parameters, such as tumor size and nodal status, with prognostic gene expression data to provide a single prognostic score. The performance of the second-generation tests in the prediction of long-term outcomes is higher compared with the first-generation tests (7).

In advanced or metastatic HR⁺/HER2⁻ breast cancer, no validated prognostic biomarker exists to date (8). Nonetheless, several clinical parameters have been associated with poor outcomes, including a higher tumor burden, a poor performance status, and aggressive features, such as visceral disease and a high number of metastasis (9–13). Recently, molecular studies in advanced HR⁺/HER2⁻ disease have revealed that the prediction analysis of microarray 50 (PAM50) gene expression identifies all the intrinsic subtypes (i.e., luminal A, luminal B, HER2-enriched, and basal-like), and these entities provide a strong prognostic value beyond clinical variables (14–17). These data suggest that, similar to early HR⁺/HER2⁻ disease, both clinical parameters and genomic data provide independent prognostic information in patients with advanced HR⁺/HER2⁻ disease.

Predicting prognosis in advanced HR⁺/HER2⁻ breast cancer might have clinical utility. For example, identification of patients with outstanding outcomes with endocrine therapy (ET) only might help deescalate cyclin-dependent kinase (CDK) 4/6 inhibition (CDK4/6i) or chemotherapy. Alternatively, identification of patients with a very poor prognosis might allow the design of clinical trials with novel therapies. Thus, a tool that identifies both groups of patients might be of value. Here, we report the development and validation of the first combined clinicogenomic prognostic biomarker in patients with advanced breast cancer treated with endocrine-based therapies (18–21).

The eligibility criteria, study design, and results of the EGF30008 trial were reported previously (18, 19). Briefly, 1,286 patients with stage III or IV postmenopausal HR⁺ breast cancer previously untreated in the metastatic setting were randomized in a blinded fashion to receive letrozole 2.5 mg/day with either lapatinib 1,500 mg/day or placebo. All patients provided written informed consent before enrollment. Patients were stratified by sites of disease (soft tissue/visceral or bone-only disease) and prior adjuvant antiestrogen therapy (<6 months since discontinuation or ≥6 months since discontinuation or no prior endocrine therapy (ET). Hormone receptor–positive was determined per the enrolling site and HER2 status was determined in a commercial laboratory in primary or metastatic sites defined as either FISH-positive, 3+ staining by IHC, or 2+ stainings by IHC and confirmed HER2 FISH-positive. The study is registered at clinicaltrials.gov (NCT00073528).

BOLERO-2 was an international, double-blind phase III trial, in which postmenopausal women with advanced, aromatase inhibitor–refractory, HR⁺/HER2⁻ breast cancer, measurable disease (or mainly lytic bone lesions), and Eastern Cooperative Oncology Group performance status (ECOG PS) of 0–2 were randomized (2:1) to receive oral everolimus (10 mg/day) or matching placebo plus oral exemestane (25 mg/day) until disease progression, unacceptable toxicity, or withdrawal of consent (20, 21). Randomization was stratified according to the presence of visceral metastases and prior sensitivity to ET. All patients provided written informed consent before enrollment. Eligibility criteria have been previously reported (20). Key exclusion criteria included a history of brain metastases and prior exposure to exemestane or mTOR inhibitors. The study is registered at clinicaltrials.gov (NCT00863655). The studies were conducted in accordance with the Declaration of Helsinki ethical guidelines and approved by ethics boards.

A section of the formalin-fixed paraffin-embedded (FFPE) breast tissue was first examined with hematoxylin and eosin staining to confirm presence of invasive tumor cells and determine the tumor area. RNA purification (Roche High Pure FFPET RNA Isolation Kit) was carried out using ≥1–3 10-μm FFPE slides for each tumor specimen; macrodissection was performed, when needed, to avoid normal breast tissue contamination. A minimum of approximately 150 ng of total RNA was used and expression of the Prediction Analysis of Microarray 50 (PAM50) genes and 5 housekeeping genes [actin beta (ACTB), mitochondrial ribosomal protein L19 (MRPL19), proteasome 26S subunit, ATPase 4 (PSMC4), ribosomal protein lateral stalk subunit P0 (RPLP0), and splicing factor 3a subunit 1 (SF3A1)] were analyzed using the nCounter platform (NanoString Technologies). Data were log base 2 transformed and normalized using 5 housekeeping genes. Samples with ≤20 counts in ≥70% of the genes were excluded.

Data that passed sample and assay quality metrics were provided in a blinded fashion to NanoString Technologies for normalization and analysis with a proprietary PAM50 algorithm. Intrinsic subtyping (luminal A, luminal B, HER2-enriched, basal-like, and normal-like) was performed using the research-based PAM50 intrinsic subtype predictor as described previously (14, 22).

The prognostic model was developed using HR⁺/HER2⁻ tumor samples from 484 patients enrolled in the EGF30008 trial who did not receive prior ET or relapse ≥6 months from tamoxifen discontinuation. Patient samples were split into training (67% of samples, n = 326) and validation (remaining 33% of samples, n = 158) sets, balancing for progression-free survival (PFS) event and treatment arm. Prognostic models of different feature sets were compared by C-index, the index of rank concordance for survival data. A maximum of 65 variables were evaluated. Feature sets included (i) PAM50 subtype assignments (n = 5); (ii) PAM50 subtype expression signatures (PAM50 scores), defined as correlations to each of the prototypic centroids (n = 5); (iii) expression of 50 individual genes that constitute the PAM50 assay; and (iv) 5 clinical variables: ECOG PS, visceral disease, number of metastasis, age, and type of biopsy used for gene expression analysis (primary versus metastatic). These feature sets were evaluated within training set (67% of samples) by Monte-Carlo cross validation (MCCV) with 100 iterations. During each iteration, training samples were split into MCCV-training (67%) and MCCV-test (33%) sets, balancing for PFS. Cox proportional hazard models were fit with ElasticNet in each iteration of MCCV training and evaluated in the MCCV test sets. Alpha and lambda were selected on the basis of the mean C-index from the MCCV test sets. Selected parameters were then used to fit a model of the complete training dataset (n = 484) for evaluation in BOLERO-2. A cutoff was defined to split patients into high- and low-risk groups. The cutoff utilized for the BOLERO-2 evaluation was derived using power calculations relating to the expected sample size of BOLERO-2 and HRs) estimated from candidate thresholds in EGF30008. The resulting cutoff was estimated to generate high-risk calls for 40% of the BOLERO-2 cohort.

Univariate and multivariable Cox proportional hazard regression analyses were performed to investigate the association of each variable with PFS and overall survival (OS). The significance level was set to a two-sided alpha of 0.05. We used R version 3.5.1. for all the statistical analyses.

Prognostic models were developed from 484 patients (37.6%) enrolled in the EGF30008 trial (Fig. 1A). The main criteria were HR⁺/HER2⁻ tumor samples and no prior ET or relapse ≥6 months from tamoxifen discontinuation. The median (range) age was 64 (31–94) years, and 260 (54%) patients had a performance status of 0, 247 (51%) had less than 3 metastatic sites, 422 (87%) had visceral disease, and 244 (50%) received lapatinib (Table 1). Subtype distribution was as follows: luminal A (52.9%), luminal B (31.4%), normal-like (11.1%), basal-like (2.7%), and HER2-enriched (1.9%). Similar to the original results, lapatinib therapy did not show a significant PFS benefit in patients with HER2-negative disease (data not shown).

The following PFS prognostic models were built and tested in the EGF30008 dataset using the following data: (i) clinical variables only; (ii) clinical variables, including PAM50 subtypes and PAM50 scores; (iii) 50 genes; (iv) PAM50 subtypes, PAM50 scores and 50 genes; and (v) clinical variables, PAM50 subtypes, PAM50 scores, and 50 genes (Supplementary Fig. S1). Among them, models that included clinical variables, PAM50 subtypes, PAM50 scores, and 50 genes yielded the highest C-index scores. This final model included a total of 21 variables and was named PAM50MET (Supplementary Table S1).

The PAM50MET variables associated with a better PFS (i.e., low PAM50MET scores) were ECOG PS of 0 (versus 1), older age, PAM50 luminal A subtype (vs. not), and high expression of the following genes: solute carrier family 39 member 6 (SCL39A6), also known as LIV-1, microtubule associated protein tau (MAPT), progesterone receptor (PGR), N-acetyltransferase 1 (NAT1), estrogen receptor 1 (ESR1), and transmembrane protein 45B (TMEM45B). The PAM50MET variables associated with a worse PFS (i.e., high PAM50MET scores) were ≥3 metastatic sites (vs. <3 metastatic sites), primary tumor (vs. metastatic tumor), PAM50 HER2-enriched subtype, PAM50 basal-like signature score, and high expression of the following genes: cyclin B1 (CCNB1), phosphoglycerate dehydrogenase (PHGDH), fibroblast growth factor receptor 4 (FGFR4), growth factor receptor bound protein 7 (GRB7), forkhead box A1 (FOXA1), NUF2 component of NDC80 kinetochore complex (NUF2), G protein–coupled receptor 160 (GPR160), and ubiquitin conjugating enzyme E2 T (UBE2T).

The prognostic value of the PAM50MET model was explored in the EGF30008 training dataset (Fig. 2). The PAM50MET-low group, defined by a predefined cutoff (i.e., −1.065), was found to be significantly associated with a better PFS compared with the PAM50MET-high group [HR = 0.37; 95% confidence interval (CI), 0.29–0.47; P < 0.0001; Fig. 2A). The median (95% CI) PFS was 22.24 (19.0–24.9) months in the PAM50MET-low group versus 9.13 (8.15–11.0) months in the PAM50MET-high group.

Similarly, the PAM50MET-low group was found to be significantly associated with a better OS versus PAM50MET-high group (HR = 0.37; 95% CI, 0.27–0.51; P < 0.0001; Fig. 2B). Finally, when the PAM50MET score was grouped into quartiles, a clear gradient in the risk of progression or death was observed (Fig. 2C and D). The median (95% CI) OS was not reached in the PAM50MET-low group versus 33.0 (28.0–40.0) months in the PAM50MET-high group. Similar results were obtained when each treatment arm (i.e., lapatinib or placebo) was evaluated separately (Supplementary Figs. S2 and S4).

The prognostic value of the PAM50MET model was tested in 261 patients enrolled in the BOLERO-2 trial (Fig. 1B; Table 2). The PAM50MET-low group, defined by the predefined cutoff, was found to be significantly associated with a better PFS versus the PAM50MET-high group (HR = 0.72; 95% CI, 0.53–0.96; P < 0.0001; Fig. 3A). The median (95% CI) PFS was 6.93 (5.57–11.0) months in the PAM50MET-low group versus 5.23 (4.2–6.8) months in the PAM50MET-high group. Similarly, the PAM50MET-low group was found to be significantly associated with a better OS compared with the PAM50MET-high group (HR = 0.49; 95% CI, 0.35–0.69; P < 0.0001; Fig. 3B). The median (95% CI) OS was 36.9 (33.4–NA) months in the PAM50MET-low group versus 23.5 (20.2–28.3) months in the PAM50MET-high group. When the PAM50MET score was grouped into quartiles, a gradient in the risk of progression or death was observed (Fig. 3C and D). Finally, similar results were obtained when each treatment arm (i.e., everolimus or placebo) was evaluated separately (Supplementary Figs. S3 and S4).

ET alone or in combination with CDK4/6i is the preferred choice of first-line treatment for pre/postmenopausal women with HR⁺/HER2⁻ breast cancer (excluding those with visceral crisis; ref. 8). The choice of ET depends on prior line of agents received in early disease, thereby highlighting the need to consider prior exposure in treatment decision-making (8). Everolimus plus aromatase inhibitor (AI) is a recommended treatment in second-line settings (20, 23). Recently, alpelisib (a PI3K inhibitor) in combination with fulvestrant has been approved for the treatment of postmenopausal patients with HR⁺/HER2⁻, PIK3CA-mutated (occur in about 40% of cases), advanced or metastatic breast cancer following progression or after an endocrine-based regimen (24). Overall, the goal is to improve survival and quality of life and delay the use of chemotherapy.

To date, very few prognostic molecular biomarkers have been reported in advanced HR⁺/HER2⁻ disease (25). In addition, prior studies have also indicated (11, 14, 15, 26) that along with molecular biomarkers such as intrinsic subtyping, other clinical variables including subtype, prior ET intake, age, ECOG PS, presence/absence of visceral metastases, sites of metastasis, number of metastatic sites, Scarff–Bloom–Richardson grade, disease-free interval, and type of tumor tissue are crucial in predicting survival outcomes in HR⁺/HER2⁻ advanced breast cancer. On the basis of this, the PAM50MET model was derived considering the above-mentioned clinical variables and PAM50 subtyping.

Overall, this is the first study to validate a combined clinicogenomic prognostic biomarker in patients with advanced HR⁺/HER2⁻ breast cancer who received endocrine-based therapies. The data demonstrated a significant benefit in PFS (the PAM50MET-low vs. PAM50MET-high group in BOLERO-2; HR = 0.72) and OS (the PAM50MET-low vs. PAM50MET-high group in BOLERO-2; HR = 0.49) outcomes. The PAM50MET score was mainly designed to predict PFS; however, based on the data observed, it was witnessed that PAM50MET was a robust predictor for OS (HR = 0.49) compared with PFS (HR = 0.72). This OS benefit can be hypothesized in a perspective that the low-risk tumors might have a biology that is likely to respond to subsequent therapies, whereas the medium-high-risk tumors might have a biology that does not respond that well to subsequent therapies.

Previous studies in early HR⁺/HER2⁻ breast cancer (27–29) support the use of multianalyte gene testing methods such as PAM50 or MammaPrint or OncotypeDX, together with clinical variables, to help predict clinical outcomes and the need of adjuvant chemotherapy (30). The clinical utility of similar biomarker approaches in advanced HR⁺/HER2⁻ breast cancer is currently unknown. However, with further validation, especially in studies with ET +/− CDK4/6 inhibitors, the PAM50MET score might be helpful in identifying patients who could be treated with ET only, for example, those with newly diagnosed with de novo metastatic disease, and patients with very poor clinical outcomes despite standard therapies and in need of novel treatment options. Tumors from patients with very poor clinical outcomes (i.e., PAM50MET-high) might have high clonal diversity and high mutational tumor burden. All of these hypotheses need further validation in other studies.

This study has few limitations. As discussed earlier, the PAM50MET score used for model development was trained using patient samples from the EGF30008 trial, which is a first-line therapy AI-naïve cohort. On the contrary, the testing data set (BOLERO-2) had patients in second-line settings. Yet, it is noteworthy that the PAM50MET score showed prognostic performance in the second-line setting as well, despite different sample sets. On the contrary, the predictive value of the PAM50MET in patients treated in other treatment settings is unknown, and the question remains if it could predict benefit to patients treated with CDK4/6 or ET. This will pave the path in testing this hypothesis in future randomized trials. It is still unknown whether it is better to perform gene expression in a metastatic biopsy or primary tumor biopsy although both the studies had approximately 80% primary tumor biopsies. Finally, the clinical utility of PAM50MET is currently unknown. Particular cut-off points and further validation is needed.

Our results support the use of the PAM50MET model to improve the classification of patients with advanced HR+/HER2-negative BC into prognostic groups, allowing for a more precise identification of progression and death risks and an improved basis for treatment decisions. The PAM50MET model for risk classification may result in a significant reduction in the use of chemotherapy in the advanced setting and may also help in the identification of potential candidates for ET only or novel therapies.

A. Prat reports grants and personal fees from Roche and personal fees from Novartis and NanoString Technologies during the conduct of the study, and personal fees from AstraZeneca, Pfizer, Daiichi Sankyo, Oncolytics Biotech, Foundation Medicine, and Guardant Health outside the submitted work. J.C. Brase reports employment and stocks with Novartis. S.R.D. Johnston reports personal fees from Eli Lilly (consultancy/advisory role; research funding to institution), Novartis (consultancy/advisory role, speaker honorarium, research funding to institution), personal fees and other from AstraZeneca (consultancy/advisory role, speaker honorarium, research funding to institution), Pfizer (consultancy/advisory role, speaker honorarium, research funding to institution), Puma Biotechnology (consultancy/advisory role, research funding to institution), and Eisai (consultancy/advisory role, speaker honorarium) outside the submitted work. E.M. Ciruelos reports personal fees from Pfizer (speaker bureau, travel grants, advisory board meetings), Roche (speaker, advisory board meetings, travel grants), Lilly (speaker, advisory board meetings), Novartis (speaker, advisory board meetings), Astra Zeneca (speaker, advisory board meetings), and MSD (advisory board meetings) outside the submitted work. J.S. Parker reports patents for US9631239B2 and EP2664679B1 issued to NanoString. No potential conflicts of interest were disclosed by the other authors.

A. Prat: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing-original draft, project administration, writing-review and editing. Y.-H. Tsai: Conceptualization, resources, data curation, software, formal analysis, validation, investigation, visualization, methodology, writing-original draft, writing-review and editing. T. Pascual: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing-original draft, project administration, writing-review and editing. L. Paré: Conceptualization, resources, data curation, software, formal analysis, supervision, validation, investigation, visualization, methodology, writing-original draft, project administration, writing-review and editing. B. Adamo: Data curation, software, formal analysis, validation, investigation, visualization, methodology, writing-original draft, writing-review and editing. M. Vidal: Data curation, software, formal analysis, validation, investigation, visualization, methodology, writing-original draft, writing-review and editing. F. Brasó-Maristany: Data curation, software, formal analysis, funding acquisition, validation, investigation, visualization, methodology, writing-original draft, writing-review and editing. P. Galván: Data curation, software, formal analysis, supervision, validation, investigation, visualization, methodology, writing-original draft, project administration, writing-review and editing. J.C. Brase: Conceptualization, resources, data curation, software, formal analysis, supervision, validation, investigation, visualization, methodology, writing-original draft, writing-review and editing. V. Rodrik-Outmezguine: Data curation, software, formal analysis, validation, investigation, visualization, methodology, writing-original draft, writing-review and editing. S. Johnston: Data curation, software, formal analysis, validation, investigation, visualization, methodology, writing-original draft, writing-review and editing. E. Ciruelos: Data curation, software, formal analysis, validation, investigation, visualization, methodology, writing-original draft, writing-review and editing. J.S. Parker: Conceptualization, resources, data curation, software, formal analysis, supervision, validation, investigation, visualization, methodology, writing-original draft, project administration, writing-review and editing.

This work was supported by Instituto de Salud Carlos III (PI16/00904 and PI19/01846. to A. Prat), AACR Career Development Awards for Translational Breast Cancer Research, the Breast Cancer Research Foundation (19–20–26-PRAT, to A. Prat), Breast Cancer Now (2018NovPCC1294, to A. Prat), RESCUER, funded by European Union's Horizon 2020 Research and Innovation Programme under Grant Agreement No. 847912, Fundació La Marató TV3 (201935–30, to A. Prat), Fundación Mutua Madrileña, and Fundación Quirón-Salud (to A. Prat), Fundación Científica Asociación Española Contra el Cáncer (Ayuda Postdoctoral AECC 2017, to F. Brasó-Maristany); and Fundación SEOM (Beca FSEOM 2018 de Formación de 2 años de estancia en Centro de Referencia en el extranjero, to T. Pascual). We thank all the patients and their family members for participating in the 2 studies. The authors thank Bhavani Yamsani, MPharm, MBA of Novartis Healthcare Pvt Ltd (Hyderabad, India), for providing medical editorial assistance in the preparation of this manuscript, in accordance with Good Publication Practice (GPP3) guidelines.

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.