Purpose:

There is currently no molecular signature in clinical use for adjuvant endocrine therapy omission in breast cancer. Given the unique trial design of SweBCG91-RT, where adjuvant endocrine and chemotherapy were largely unadministered, we sought to evaluate the potential of transcriptomic profiling for identifying patients who may be spared adjuvant endocrine therapy.

Experimental Design:

We performed a whole-transcriptome analysis of SweBCG91-RT, a randomized phase III trial of ± radiotherapy after breast-conserving surgery for node-negative stage I–IIA breast cancer. Ninety-two percent of patients were untreated by both adjuvant endocrine therapy and chemotherapy. We calculated 15 transcriptomic signatures from the literature and combined them into an average genomic risk, which was further used to derive a novel 141-gene signature (MET141). All signatures were then independently examined in SweBCG91-RT and in the publicly available METABRIC cohort.

Results:

In SweBCG91-RT, 454 patients were node-negative, postmenopausal, and systemically untreated with ER-positive, HER2-negative cancers, which constitutes a low-risk subgroup and potential candidates for therapy omission. Most transcriptomic signatures were highly prognostic for distant metastasis, but considerable discordance was observed on the individual patient level. Within the MET141 low-risk subgroup (lowest 25th percentile of scores), 95% of patients were free of metastasis at 15 years, even in the absence of adjuvant endocrine therapy. In a clinically low-risk subgroup of the METABRIC cohort not treated with systemic therapy, no breast cancer death occurred among the MET141 low-risk patients.

Conclusions:

Transcriptomic profiling identifies patients with an excellent outcome without any systemic adjuvant therapy in clinically low-risk patients of the SweBCG91-RT and METABRIC cohorts.

Translational Relevance

Some women with primary breast cancer do not require additional endocrine therapy after breast-conserving surgery, but no tests are in use to find this low-risk group of women. We performed a transcriptomic analysis of 765 patients of the SweBCG91-RT trial, of whom 454 were node-negative, postmenopausal, and systemically untreated with ER-positive, HER2-negative cancers. We tested 15 previously published signatures and showed that most perform well in identifying women with very low risk of recurrence. However, there was a substantial intersignature variation in risk classification, and we therefore combined the signatures into an average genomic risk and an associated novel signature (MET141). MET141 could identify a low-risk group of node-negative, postmenopausal, nonsystemically treated patients with ER+ and HER2-negative tumors, of which 95% were free of metastasis at 15 years. These results indicate that transcriptomic profiling may be used to find women who may be spared endocrine treatment.

Treatment of primary breast cancer is becoming more and more individualized and has entered the era of precision medicine. Due to increased public awareness and intensified screening programs, the proportion of low-risk tumors has increased with a corresponding risk of overtreatment (1). Thus, in addition to escalating treatment for patients with high-risk breast cancers, current guidelines focus on deescalating treatment in low-risk patients (2). Although gene signatures assessing recurrence risk have been successful at identifying patient subgroups in whom adjuvant chemotherapy can be safely omitted (3–5), there are no tests currently in clinical guidelines to identify patients who may omit endocrine therapy (2). Adjuvant endocrine therapy reduces the risk of breast cancer death in patients with estrogen receptor–positive (ER+) disease by around one-third (6), which can be further reduced by using aromatase inhibitors in postmenopausal patients (7). However, endocrine therapy may have substantial side effects, which is reflected in an adherence rate between 50% and 80% (8), and most patients with node-negative disease will not suffer a recurrence even without adjuvant systemic therapy (6). Thus, developing tools to safely omit endocrine therapy among patients with ER+ cancers is highly desirable.

One approach to personalizing therapy is to consider relative treatment effects constant over subgroups and identify patients at low risk of recurrences in the absence of the treatment in question (9). The PAM50 risk of recurrence score was shown to identify a subgroup of patients with node-positive hormone-receptor–positive tumors treated with endocrine therapy but not chemotherapy with a 10-year metastasis risk of 6.6%, suggesting that patients in this subgroup may be spared chemotherapy (3). Among women with high clinical risk but low 70-gene scores of the MINDACT trial, the five-year metastasis-free survival for those that did not receive chemotherapy was similarly high, at 94.7% (5). Furthermore, other studies have focused on identifying patients at low risk of recurrence despite not receiving any adjuvant systemic therapy. A clinically low-risk subgroup of patients with no adjuvant treatment of the Oslo1 trial with low PAM50 risk of recurrence scores had a 15-year breast cancer–specific survival of 96.3% (10). Similarly, the 70-gene signature was recently shown to identify an ultra-low-risk group of patients in the STO-3 trial with a breast cancer–specific survival rate of 94% at 20 years in the absence of both endocrine therapy and chemotherapy (11).

When considering the use of baseline risk for gene-expression tests, an emerging problem is the substantial discordance in results for an individual patient. Indeed, a recent study found the agreement of five common gene-expression tests to be modest, with 39% of patients classified uniformly as low risk by all tests, whereas individual tests predicted 61% to 82% to be low risk (12). Other barriers for identifying patients for whom adjuvant endocrine therapy can be safely withheld include the lack of studies in which patients were not treated with endocrine therapy and lack of studies with long follow-up. As ER+ breast cancer continues to recur and cause death at a relatively consistent rate over 15 years after stopping endocrine therapy, studies with long follow-up are necessary to identify patients who may experience late recurrences (13). Thus, in order to evaluate risk stratification tools for endocrine therapy there is a need for large, well-defined cohorts of patients who were not treated with adjuvant systemic therapy, have long-term follow-up, and in whom several gene-expression signatures can be compared.

To that end, we examined the transcriptome of 765 early-stage breast cancer patients from the SweBCG91-RT trial, a trial randomizing node-negative stage I–IIA breast cancer patients undergoing breast-conserving surgery to ± adjuvant whole breast radiotherapy (14). The vast majority (92%) of patients in the trial were systemically untreated in the adjuvant setting, and 454 patients were ER+, HER2-negative (HER2), postmenopausal, and did not receive adjuvant systemic therapy, making it an ideal data set to study recurrence risk in the absence of adjuvant systemic therapy. We calculated gene-expression signatures for 15 previously published signatures and aimed to evaluate the potential of transcriptomic profiling in identifying patients at such low risk of metastasis that adjuvant endocrine therapy can be safely omitted.

SweBCG91-RT patients

We analyzed gene-expression data of the SweBCG91-RT trial, the details of which have been previously described (14–16). Briefly, the trial randomized 1,178 node-negative, early-stage breast cancer patients undergoing breast-conserving surgery to adjuvant whole breast radiotherapy or no radiotherapy. As systemic adjuvant therapy was administered according to regional guidelines at the time, it was sparsely provided, with only 7% and 2% of patients in the original trial receiving endocrine therapy and chemotherapy, respectively (15). Subtyping was performed using IHC as detailed previously (14). The primary endpoint of this analysis was distant recurrence-free interval (i.e., time to metastasis), defined from the time of surgery until the time of metastasis, last follow-up or death, with death as a competing event (17). Patients suffering a contralateral breast cancer or another primary cancer were not censored, as recommended (18). The data for the metastasis endpoint were collected from patient chart review, and the median follow-up time was 15.1 years for patients free from event. Additional follow-up was derived from the Swedish cause of death registry with a median follow-up time of 20.0 years for patients alive at censoring, and we present cumulative incidence of breast cancer death, with death from other causes as competing event, as Supplementary Information. The trial and follow-up study were conducted in accordance with the declaration of Helsinki and were approved by the Lund University Regional Ethical Review Board (approval numbers 2010/127 and 2015/548). Informed oral consent was obtained from all patients, which was determined appropriate and approved by the Ethical Review Board for the original trial and for this gene-expression study.

Gene-expression analysis

Formalin-fixed paraffin-embedded tissue was available for 922 of the original 1,178 patients in the trial (Supplementary Fig. S1). RNA extraction and microarray hybridization were performed in a Clinical Laboratory Improvement Amendments certified laboratory (Decipher Biosciences). Tumors were profiled with the GeneChip Human Exon 1.0 ST microarray (Thermo Fisher) and 765 tumors passed quality control of RNA, cDNA, and microarray analysis (Gene-Expression Omnibus GSE119295). Gene-expression data were normalized using Single Channel Array Normalization (19).

Publicly available METABRIC data

We also examined gene signature scores in the Molecular Taxonomy of Breast Cancer International Consortium (METABRIC) cohort. Publicly available clinical and expression data based on the Illumina Human v3 array were downloaded from cBioPortal. Out of the 1,904 patients with microarray expression data, 104 patients were postmenopausal, treated with breast-conserving surgery, with ER+, human epidermal growth factor receptor 2 negative (HER2) tumors, complete breast cancer–specific death information, and were not treated with endocrine or chemotherapy. Nearly all were lymph node negative (20). This low-risk systemic treatment naïve group was included for analysis in this study. The median follow-up time for this low-risk subgroup was 18.1 years for patients alive at censoring.

Data analysis

Statistical analyses were performed using R (3.5.2). We performed a literature review and identified 15 previously published gene-expression signatures specific to breast cancer risk with published equations or algorithms for calculation (21–35). Most were created to prognosticate for the distant recurrence endpoint, although a few (PIK3CAGS and TAMR13) were designed for tamoxifen sensitivity. The surrogate scores of these previously published gene-expression signatures were calculated using published algorithms as described below. Cumulative incidences of metastasis or death from breast cancer were computed with a competing risks approach using the cmprsk package (36), and 95% confidence intervals were computed as previously described (37). For a direct and unbiased comparison of how the different signatures perform, the patients were grouped by score quartiles. We further examined rates of metastasis or death from breast cancer for patients with the lowest quartile of risk scores, hypothesizing that these patients may be candidates for therapy omission, although aware that this may not directly represent clinical cutoffs used for the signatures. Cause-specific Cox proportional hazards regression was used to contrast the differences in hazards between patients with high and low signature scores, and P values were computed with the Wald test. Each continuous risk score was standardized by dividing the score by its standard deviation in order to create comparable hazard ratios across signatures, otherwise signatures with smaller ranges of values would have disproportionately higher hazard ratios, and the hazard ratios would not be comparable. Proportional hazards were checked graphically and by the Schoenfeld test (38). For most signatures, the hazard ratio (HR) was larger during the first years of follow-up, and we therefore limited this analysis to 10 years. However, a trend was still observed with larger hazard ratios years 0–5 than 5–10, and the presented HRs should thus be interpreted as the mean over 10 years. To compare how classification of low-risk patients differs by signature, we identified patients within the lowest quartile of risk scores of an individual signature and calculated the proportion of those patients also classified in the lowest quartile of each other signature. We then computed the mean proportion, excluding the signature of interest (in which the proportion is 1). We refer to this as the “low-risk classification agreement.” In addition, the agreement between signatures split by quartiles was tested by calculating Cohen's kappa. We followed REMARK guidelines for reporting of this study (39). Adjustment of P values for multiple testing correction were performed using the Benjamini–Hochberg false discovery rate (FDR) method, where applicable (40).

Estimation of time-dependent area under the curve

Estimation of time-dependent area under the curve (AUC) was calculated using the R survivalROC package (version 1.0.3) (41). 95% confidence intervals for time-dependent AUC estimates were bootstrapped using 1,000 bootstrap samples.

Pathway analysis

To assess biological pathways overrepresented in lists of genes, we used the Panther statistical overrepresentation test (version 13.0, pantherdb.org; ref. 42) using Fisher exact test with Benjamini–Hochberg FDR correction as the test type, and Panther GO-Slim Biological Process gene lists as the annotation data set. As a secondary method, we also used Reactome Analysis Tools (reactome.org; refs. 43, 44) with the “project to human” option. The Reactome genome-wide overview of the pathway analysis visualizes the enrichment analysis by organizing Reactome pathways in a hierarchy. The top-level pathway is represented as the center of a circular “burst” and each next level lower on the pathway hierarchy is represented by a step away from the center. Pathways overrepresented in the input data set are represented in yellow and pathways not significantly overrepresented are represented in gray. For both methods, lists of official gene symbols were entered. Significant enrichment of a pathway was defined as FDR < 0.05.

Computation of previously published breast cancer risk scores

Previously published breast cancer risk scores were developed on a variety of platforms. We applied gene-expression data from microarrays to genomic signature equations to calculate surrogate continuous risk scores. The following risk scores were calculated according to their equations as published, using the genefu package (version 2.6.0; ref. 45) in R (version 3.3.2): OncotypeDx-like (21), Endopredict-like (22), Genomic Grade Index-like (23), PAM50ROR-like (24), Gene70-like (46), GeniusM3-like (26), TAMR-like (27), Gene76-like (28), and the PIK3CAGS-like risk score (29). For signatures that are based on probes from specific microarrays, the genefu annotations to Entrez gene identifiers were used to map probes to the appropriate gene on the microarray platform. For the few genes not available on microarray, the term (coefficient and gene-expression value) of that gene was omitted from the signature equation. The genefu functions used are listed in Supplementary Table S1.

Celera-like risk score

Risk scores were computed by calculating the sum of the expression of 14 genes, as previously described (30).

ExagenBC-like ER+ risk score

Risk scores were computed based on the following equation: R = 0.128*CYP24 – 0.173*PDCD6IP + 0.183*BIRC5, as previously described (31).

Mammostrat-like risk score

Risk scores were computed based on the following equation: R = 1.54*SLC7A5 + 1.12*TP53 + 1.06*NDRG1 + 0.72*HTF9C + 0.5* CEACAM5, as previously described (32).

MGI-like risk score

Risk scores were computed by normalizing the expression levels for each of the five genes in the score to have a mean of 0 and a standard deviation of 1, then combined into a single score as the first principal component, as previously described (33).

Toronto 2017-like risk score

Risk scores were computed by calculating the linear equation involving gene expression and coefficients of 95 genes, as previously described (34).

Two-gene ratio-like risk score

Risk scores were computed by subtracting the expression of IL17RB from the expression of HOXB13, as previously described (35).

Average genomic risk

To calculate average genomic risk, each of the 15 signature scores was scaled from 0 to 1 within the cohort, and then the mean was computed. The scaling was necessary to prevent signatures with larger ranges of values to be overweighted in the calculation of the average risk.

MET141

We performed a literature search to identify publicly available gene-expression data sets with metastasis available as an endpoint. These publicly available breast cancer data sets were Servant (GSE30682), Kao (GSE20685), Wang (GSE2034), Symmans (GSE17705), and van de Vijver (downloaded from http://smd.princeton.edu/; refs. 25, 28, 46–48). We sought a wide range of breast cancer patients to be able to capture underlying breast cancer risk. All patients in these data sets were used for analysis and they represent breast cancer patients with a range of clinical risk factors and treatment. Briefly, the Servant cohort included 343 patients with early-stage breast cancer all treated with breast-conserving surgery and postoperative radiotherapy, with a mix of adjuvant systemic treatment. The Kao cohort included 327 patients randomly selected from the institutional tumor bank with a range of low-risk and high-risk clinical risk factors. The Wang cohort included 286 lymph node–negative patients who did not receive systemic neoadjuvant or adjuvant therapy. The Symmans cohort included 508 patients with HER2 tumors, treated with chemotherapy. The van de Vijver cohort included 295 patients treated with mastectomy or breast-conserving surgery, with a mix of adjuvant treatment (25, 28, 47–49). For each data set, probes were converted to gene symbols, and the subset of genes in common between the five data sets was identified (10,990 genes). The 15 previously published signatures and average genomic risk were calculated for each patient in these five cohorts. To assess genes to include in a new signature, we removed genes in common with genes from the previously published signatures, and using the remaining 10,315 genes, correlated each gene to the average genomic risk within each cohort. Genes with a Spearman correlation coefficient > 0.4 or <−0.4 to average genomic risk in all five cohorts were retained, resulting in 141 total genes: 89 positively correlated genes and 52 negatively correlated genes (Supplementary Table S2). The correlation coefficient value was initially varied from 0.3, 0.4, and 0.5. We found that using cutoffs of 0.3 or 0.5, the signature was not prognostic in all five training cohorts. The final MET141 score is the average expression of negatively correlated genes subtracted from the average expression of positively correlated genes.

SweBCG91-RT cohort characteristics

The SweBCG91-RT cohort was enriched for ER+ and HER2 tumors. Ninety-two percent of patients were systemic treatment naïve and did not receive adjuvant endocrine therapy or chemotherapy (Table 1). We obtained gene-expression data from 765 patients (Supplementary Fig. S1), of which 85% were free of metastasis event at 15 years. In this gene-expression analysis of the SweBCG91-RT cohort, risk scores from 15 previously published gene-expression signatures were calculated and assessed for prognostic potential for metastasis and death from breast cancer. Thirteen of the 15 calculated scores from previously published signatures were prognostic (P < 0.05) in the full SweBCG91RT cohort with respect to metastasis (Supplementary Fig. S2), with similar results for death from breast cancer (Supplementary Fig. S3).

Table 1.

SweBCG91-RT patient characteristics.

All patientsER+, HER2, postmenopausal, no systemic treatment
Number of patients 765 454 
Age at surgery 
Median (range) 59 (31–78) 63 (39–78) 
 ≤39 19 (3%) 1 (0%) 
 40–49 137 (18%) 16 (4%) 
 50–59 234 (31%) 151 (33%) 
 60–69 284 (37%) 210 (46%) 
 ≥70 91 (12%) 76 (17%) 
Menopausal status 
 Premenopausal 152 (20%) 0 (0%) 
 Postmenopausal 592 (80%) 454 (100%) 
 Missing 21 
Histologic grade 
 1 105 (14%) 73 (16%) 
 2 457 (61%) 312 (70%) 
 3 191 (25%) 61 (14%) 
 Missing 12 
Tumor size (mm) 
Median (range) 12 (1–40) 11 (1–30) 
 ≤10 274 (36%) 198 (43%) 
 11–20 415 (55%) 243 (54%) 
 21–30 70 (9%) 10 (2%) 
 ≥31 1 (0%) 0 (0%) 
 Missing 
Estrogen receptor status (≥1% by IHC) 
 Negative 89 (12%) 0 (0%) 
 Positive 672 (88%) 454 (100%) 
 Missing 
Progesterone receptor status (≥20% by IHC) 
 Negative 206 (27%) 90 (20%) 
 Positive 555 (73%) 364 (80%) 
 Missing 
HER2 status by IHC and FISH 
 Negative 702 (93%) 454 (100%) 
 Positive 54 (7%) 0 (0%) 
 Missing 
Subtype by IHC 
 Luminal A 421 (56%) 287 (63%) 
 Luminal B (HER2216 (29%) 167 (37%) 
 HER2+ 54 (7%) 0 (0%) 
 Triple-negative 65 (9%) 0 (0%) 
 Missing 
Adjuvant endocrine therapy 
 No 710 (93%) 454 (100%) 
 Yes 55 (7%) 0 (0%) 
Adjuvant chemotherapy 
 No 755 (99%) 454 (100%) 
 Yes 10 (1%) 0 (0%) 
Adjuvant radiotherapy 
 No 403 (53%) 227 (50%) 
 Yes 362 (47%) 227 (50%) 
Distant metastasis 
 No 658 (86%) 402 (89%) 
 Yes 107 (14%) 52 (12%) 
Died from breast cancer 
 No 628 (82%) 373 (82%) 
 Yes 137 (18%) 81 (18%) 
All patientsER+, HER2, postmenopausal, no systemic treatment
Number of patients 765 454 
Age at surgery 
Median (range) 59 (31–78) 63 (39–78) 
 ≤39 19 (3%) 1 (0%) 
 40–49 137 (18%) 16 (4%) 
 50–59 234 (31%) 151 (33%) 
 60–69 284 (37%) 210 (46%) 
 ≥70 91 (12%) 76 (17%) 
Menopausal status 
 Premenopausal 152 (20%) 0 (0%) 
 Postmenopausal 592 (80%) 454 (100%) 
 Missing 21 
Histologic grade 
 1 105 (14%) 73 (16%) 
 2 457 (61%) 312 (70%) 
 3 191 (25%) 61 (14%) 
 Missing 12 
Tumor size (mm) 
Median (range) 12 (1–40) 11 (1–30) 
 ≤10 274 (36%) 198 (43%) 
 11–20 415 (55%) 243 (54%) 
 21–30 70 (9%) 10 (2%) 
 ≥31 1 (0%) 0 (0%) 
 Missing 
Estrogen receptor status (≥1% by IHC) 
 Negative 89 (12%) 0 (0%) 
 Positive 672 (88%) 454 (100%) 
 Missing 
Progesterone receptor status (≥20% by IHC) 
 Negative 206 (27%) 90 (20%) 
 Positive 555 (73%) 364 (80%) 
 Missing 
HER2 status by IHC and FISH 
 Negative 702 (93%) 454 (100%) 
 Positive 54 (7%) 0 (0%) 
 Missing 
Subtype by IHC 
 Luminal A 421 (56%) 287 (63%) 
 Luminal B (HER2216 (29%) 167 (37%) 
 HER2+ 54 (7%) 0 (0%) 
 Triple-negative 65 (9%) 0 (0%) 
 Missing 
Adjuvant endocrine therapy 
 No 710 (93%) 454 (100%) 
 Yes 55 (7%) 0 (0%) 
Adjuvant chemotherapy 
 No 755 (99%) 454 (100%) 
 Yes 10 (1%) 0 (0%) 
Adjuvant radiotherapy 
 No 403 (53%) 227 (50%) 
 Yes 362 (47%) 227 (50%) 
Distant metastasis 
 No 658 (86%) 402 (89%) 
 Yes 107 (14%) 52 (12%) 
Died from breast cancer 
 No 628 (82%) 373 (82%) 
 Yes 137 (18%) 81 (18%) 

Performance of calculated scores from 15 previously published signatures in potential candidates for omission of systemic adjuvant treatment

To focus on patients who could be clinical candidates for omission of systemic adjuvant treatment, we selected patients with ER+, HER2 tumors who were postmenopausal, node-negative, and did not receive any systemic adjuvant treatment (N = 454, 59% of the profiled cohort). In this low-risk subgroup, 88% of patients were free of metastasis at 15 years.

Twelve of the 15 signatures were significantly associated with metastasis (P < 0.05), with scaled 10 year HRs of 1.5 to 2.4 (Fig. 1A and B). The same set of signatures were also prognostic for death from breast cancer (Supplementary Fig. S4). As risk of late recurrences is a major concern for breast cancer patients, we analyzed the performance of the different signatures by calculating the AUC at different time points. For most signatures, there was a drop in prognostic ability over time (Supplementary Fig. S5), with an average AUC of 0.73, 0.66, and 0.60 at 5, 10, and 15 years, respectively.

Figure 1.

Performance of previously published signatures, AGR and a novel signature, MET141, in 454 node-negative, postmenopausal, and systemically untreated patients of the SweBCG91-RT trial with ER+, HER2 cancers. A, Forest plot depicting standardized HRs for each of the 15 previously published gene signatures, the AGR derived as a mean of all signatures, and a novel signature MET141, for the 454 postmenopausal and systemically untreated patients with ER+, HER2 cancers, with associated P values from the Cox proportional hazards model. Continuous risk scores were divided by the standard deviation to directly compare HRs between scores with differently distributed values, and the Cox analysis is limited to 10 years. Results are shown for the distant metastasis endpoint. B, Cumulative incidence of distant metastasis in the 454 node-negative postmenopausal patients of the SweBCG91-RT cohort with ER+, HER2 cancers who did not receive systemic therapy, for each of the 15 previously published gene signatures, AGR, and MET141.

Figure 1.

Performance of previously published signatures, AGR and a novel signature, MET141, in 454 node-negative, postmenopausal, and systemically untreated patients of the SweBCG91-RT trial with ER+, HER2 cancers. A, Forest plot depicting standardized HRs for each of the 15 previously published gene signatures, the AGR derived as a mean of all signatures, and a novel signature MET141, for the 454 postmenopausal and systemically untreated patients with ER+, HER2 cancers, with associated P values from the Cox proportional hazards model. Continuous risk scores were divided by the standard deviation to directly compare HRs between scores with differently distributed values, and the Cox analysis is limited to 10 years. Results are shown for the distant metastasis endpoint. B, Cumulative incidence of distant metastasis in the 454 node-negative postmenopausal patients of the SweBCG91-RT cohort with ER+, HER2 cancers who did not receive systemic therapy, for each of the 15 previously published gene signatures, AGR, and MET141.

Close modal

Most of the continuous risk scores were highly correlated to each other (Fig. 2A). To further visualize agreement of signatures, we created barplots where each row depicts the calculated scores and clinical information for an individual patient. All patients are sorted by the average of the 15 previously published signatures. Despite high correlation of signatures, there was considerable disagreement across signatures for an individual patient (Fig. 2B). When comparing risk scores with subtype, Ki67 and histologic grade, grade 3 and the Luminal B subtype had higher risk compared with grades 1 to 2 and the Luminal A subtype. In addition, high Ki67 was strongly correlated with higher risk scores (Fig. 2B; Supplementary Fig. S6A and S6B). Further, patients developing early recurrences tended to be classified as higher risk by most continuous risk scores (Fig. 2B), and signatures were better at identifying early recurrences, as shown by a higher AUC for all prognostic signatures for early recurrences (5 years) than for late recurrences (15 years; Supplementary Fig. S5).

Figure 2.

Comparison of previously published signatures in 454 low-risk patients of SweBCG91-RT. A, Pearson correlation and hierarchical clustering for the gene signatures. A moderate to high correlation is seen for most signatures developed in or for breast cancer patients with ER+ cancer. B, Comparison of the previously published signatures on their classification of individual patients. Each row represents an individual patient, and samples are ordered by AGR. Bar plots are colored to indicate what quartile the patient was scored per signature, with red indicating that the patient was scored with highest risk (top 25th percentile) and blue indicating that the patient was scored with lowest risk (bottom 25th percentile). Histologic grade, time to metastasis, and subtype based on IHC scores are also displayed for comparison. C, Concordance of the signatures in classifying which patients are in the lowest quartile of risk. Bar plots show the proportion of patients classified in the lowest quartile with the title signature, which was also in the lowest quartile of each other signature. This analysis is performed for the 454 postmenopausal and systemically untreated patients with ER+, HER2 cancers in the SweBCG91-RT cohort.

Figure 2.

Comparison of previously published signatures in 454 low-risk patients of SweBCG91-RT. A, Pearson correlation and hierarchical clustering for the gene signatures. A moderate to high correlation is seen for most signatures developed in or for breast cancer patients with ER+ cancer. B, Comparison of the previously published signatures on their classification of individual patients. Each row represents an individual patient, and samples are ordered by AGR. Bar plots are colored to indicate what quartile the patient was scored per signature, with red indicating that the patient was scored with highest risk (top 25th percentile) and blue indicating that the patient was scored with lowest risk (bottom 25th percentile). Histologic grade, time to metastasis, and subtype based on IHC scores are also displayed for comparison. C, Concordance of the signatures in classifying which patients are in the lowest quartile of risk. Bar plots show the proportion of patients classified in the lowest quartile with the title signature, which was also in the lowest quartile of each other signature. This analysis is performed for the 454 postmenopausal and systemically untreated patients with ER+, HER2 cancers in the SweBCG91-RT cohort.

Close modal

Based on the results analyzing rate of metastasis for patients grouped by score quartiles, we hypothesized that the lowest score quartile could be candidates for omission of therapy. To further evaluate the concordance of the 15 signatures for identifying these low-risk patients, we calculated the low-risk classification agreement, which quantifies the mean proportion of patients classified in the lowest quartile of risk by one signature also classified in the lowest quartile of risk by the other signatures. Mean classification agreement ranged from 27% to 51% (Fig. 2C). Similarly, analysis of agreement with Cohen's kappa showed none to moderate agreement (Supplementary Table S3).

Average genomic risk

In an effort to increase the stability of the prognostication, we calculated the average genomic risk (AGR) as the mean of the 15 signatures scores. The prognostic performance of AGR was in line with the most prognostic individual genomic signatures [HR = 2.1 (1.6–2.7), P < 0.001 for metastasis in the low-risk cohort; Fig. 1A and B]. Furthermore, the AGR identified a very low-risk population of patients within the ER+, HER2, postmenopausal, node-negative, and systemically untreated subgroup, as patients with the lowest quartile of AGR scores (N = 114, 25% of the subgroup) had no distant metastatic event within the first 10 years. Notably, the proportion of patients free of metastasis at 15 years was 95% (95% CI, 88%–98%; Fig. 1B).

Signature comparison and related 141-gene signature

Because many signatures were significantly associated with time to metastasis, we performed an assessment of genes shared between signatures, finding that up to 100% of genes in one signature (the MGI signature, comprised of five genes) could be found in another (Supplementary Table S4). When removing the Toronto 2017 signature from this analysis, as it had been derived using gene lists from many of the signatures included in this work, and the MGI signature, which has a small total number of genes, we found that at most 69% of genes in one signature were in common with others. Enrichment analysis for the published signatures showed that cell-cycle and metabolic pathways were significantly and highly enriched in these signature gene lists (FDR < 0.05; Supplementary Table S5). We then investigated if a signature that did not heavily share the specific genes found in these previously published signatures could still be prognostic in this data set. To that end, we derived a signature in five publicly available cohorts by identifying genes highly correlated with AGR but excluding overlapping genes with previous signatures. This 141-gene signature (MET141; Supplementary Table S2) was then independently validated in SweBCG91-RT, with a similar performance as the AGR: 95% (95% CI, 88%–98%) free of metastasis at 15 years for the lowest risk quartile in the subgroup (Fig. 1B). Gene network analysis of the AGR, comprised of the genes from the 15 previously published signatures, and MET141 gene lists suggested that both were enriched in similar gene sets with a focus on cell-cycle control, DNA replication, transcription, and extracellular matrix organization (Fig. 3A and B; Supplementary Table S6).

Figure 3.

Reactome pathway analysis. Reactome analysis pathway plots that indicate that cell-cycle, DNA replication, and gene transcription pathways are overexpressed in the gene lists for previously published signatures (A) and for the MET141 signature (B). The analysis shows that MET141 captures largely the same pathways as the previous signatures.

Figure 3.

Reactome pathway analysis. Reactome analysis pathway plots that indicate that cell-cycle, DNA replication, and gene transcription pathways are overexpressed in the gene lists for previously published signatures (A) and for the MET141 signature (B). The analysis shows that MET141 captures largely the same pathways as the previous signatures.

Close modal

Performance of calculated scores in the METABRIC cohort

We further examined if these gene signatures could identify low-risk patients who may not require adjuvant system therapy in data from METABRIC, a cohort with breast cancer–specific mortality median follow-up time of 18.1 years in patients alive at censoring. The METABRIC cohort has 1904 samples linked to microarray gene-expression data, 104 of which were from postmenopausal breast cancer patients with ER+, HER2 cancers, treated with breast-conserving surgery but no adjuvant chemotherapy or endocrine therapy, and nearly all were node negative (Table 2). In this low-risk subgroup, 83% of patients were free of breast cancer–specific death at 15 years. We calculated the aforementioned 17 signatures. Although these signatures scores were based on a different microarray platform, the majority of signatures (15/17) were able to identify a very low risk group of patients in METABRIC with low rates of breast cancer–specific death in patients with lowest 25th percentile of scores (Fig. 4). When classified by MET41, no breast cancer death occurred among patients with the lowest quartile of risk.

Table 2.

METABRIC patient characteristics.

All patientsER+, HER2, postmenopausal, treated with BCS but no systemic treatment
Number of patients 1,904 104 
Age at surgery 
 Median (range) 61.8 (21.9–96.3) 63.2 (50–87.3) 
 ≤39 116 (6%) 0 (0%) 
 40–49 295 (15%) 0 (0%) 
 50–59 431 (23%) 38 (37%) 
 60–69 552 (29%) 38 (37%) 
 ≥70 510 (27%) 28 (27%) 
Menopause status 
 Premenopausal 411 (22%) 104 (100%) 
 Postmenopausal 1,493 (78%) 0 (0%) 
Histologic grade 
 1 165 (9%) 19 (18%) 
 2 740 (39%) 60 (58%) 
 3 927 (49%) 19 (18%) 
 NA 72 (4%) 6 (6%) 
Tumor size (mm) 
 Median (range) 23 (0–182) 17 (10–43) 
 ≤10 80 (4%) 4 (4%) 
 11–20 752 (39%) 76 (73%) 
 21–30 650 (34%) 22 (21%) 
 ≥31 404 (21%) 2 (2%) 
 Missing 18 (1%) 0 (0%) 
Estrogen receptor status 
 Negative 445 (23%) 0 (0%) 
 Positive 1,459 (77%) 104 (100%) 
Progesterone receptor status 
 Negative 895 (47%) 24 (23%) 
 Positive 1,009 (53%) 80 (77%) 
HER2 status 
 Negative 1,668 (88%) 104 (100%) 
 Positive 236 (12%) 0 (0%) 
Subtype 
 ER/HER2 290 (15%) 3 (3%) 
 ER+/HER2 high proliferation 603 (32%) 35 (34%) 
 ER+/HER2 low proliferation 619 (33%) 55 (53%) 
 HER2+ 188 (10%) 2 (2%) 
 Missing 204 (11%) 9 (9%) 
Surgery type 
 Breast-conserving surgery 755 (40%) 104 (100%) 
 Mastectomy 1,127 (59%) 0 (0%) 
 Missing 22 (1%) 0 (0%) 
Adjuvant endocrine therapy 
 No 730 (38%) 104 (100%) 
 Yes 1,174 (62%) 0 (0%) 
Adjuvant chemotherapy 
 No 1,508 (79%) 104 (100%) 
 Yes 396 (21%) 0 (0%) 
Adjuvant radiotherapy 
 No 767 (40%) 15 (14%) 
 Yes 1,137 (60%) 89 (86%) 
Died from breast cancer 
 No 1,281 (67%) 82 (79%) 
 Yes 622 (33%) 22 (21%) 
 Missing 1 (0%) 0 (0%) 
All patientsER+, HER2, postmenopausal, treated with BCS but no systemic treatment
Number of patients 1,904 104 
Age at surgery 
 Median (range) 61.8 (21.9–96.3) 63.2 (50–87.3) 
 ≤39 116 (6%) 0 (0%) 
 40–49 295 (15%) 0 (0%) 
 50–59 431 (23%) 38 (37%) 
 60–69 552 (29%) 38 (37%) 
 ≥70 510 (27%) 28 (27%) 
Menopause status 
 Premenopausal 411 (22%) 104 (100%) 
 Postmenopausal 1,493 (78%) 0 (0%) 
Histologic grade 
 1 165 (9%) 19 (18%) 
 2 740 (39%) 60 (58%) 
 3 927 (49%) 19 (18%) 
 NA 72 (4%) 6 (6%) 
Tumor size (mm) 
 Median (range) 23 (0–182) 17 (10–43) 
 ≤10 80 (4%) 4 (4%) 
 11–20 752 (39%) 76 (73%) 
 21–30 650 (34%) 22 (21%) 
 ≥31 404 (21%) 2 (2%) 
 Missing 18 (1%) 0 (0%) 
Estrogen receptor status 
 Negative 445 (23%) 0 (0%) 
 Positive 1,459 (77%) 104 (100%) 
Progesterone receptor status 
 Negative 895 (47%) 24 (23%) 
 Positive 1,009 (53%) 80 (77%) 
HER2 status 
 Negative 1,668 (88%) 104 (100%) 
 Positive 236 (12%) 0 (0%) 
Subtype 
 ER/HER2 290 (15%) 3 (3%) 
 ER+/HER2 high proliferation 603 (32%) 35 (34%) 
 ER+/HER2 low proliferation 619 (33%) 55 (53%) 
 HER2+ 188 (10%) 2 (2%) 
 Missing 204 (11%) 9 (9%) 
Surgery type 
 Breast-conserving surgery 755 (40%) 104 (100%) 
 Mastectomy 1,127 (59%) 0 (0%) 
 Missing 22 (1%) 0 (0%) 
Adjuvant endocrine therapy 
 No 730 (38%) 104 (100%) 
 Yes 1,174 (62%) 0 (0%) 
Adjuvant chemotherapy 
 No 1,508 (79%) 104 (100%) 
 Yes 396 (21%) 0 (0%) 
Adjuvant radiotherapy 
 No 767 (40%) 15 (14%) 
 Yes 1,137 (60%) 89 (86%) 
Died from breast cancer 
 No 1,281 (67%) 82 (79%) 
 Yes 622 (33%) 22 (21%) 
 Missing 1 (0%) 0 (0%) 
Figure 4.

Performance of previously published signatures, AGR, and the novel signature MET141, in systemically untreated and clinically low-risk patients in the METABRIC cohort. A, Forest plot depicting standardized HRs for each of the 15 previously published gene signatures, the AGR, and the novel signature MET141, in the postmenopausal and systemically untreated patients with ER+, HER2 cancers of the METABRIC cohort, where nearly all were node negative. P values are from the Cox proportional hazards model. Continuous risk scores were divided by the standard deviation to directly compare HRs between scores with differently distributed values, and the Cox analysis is limited to 10 years. Results are shown for endpoint breast cancer death. B, Cumulative incidence of breast cancer death in the postmenopausal patients of the METABRIC cohort with ER+, HER2 cancers who did not receive systemic therapy, for each of the 15 previously published gene signatures, AGR, and MET141.

Figure 4.

Performance of previously published signatures, AGR, and the novel signature MET141, in systemically untreated and clinically low-risk patients in the METABRIC cohort. A, Forest plot depicting standardized HRs for each of the 15 previously published gene signatures, the AGR, and the novel signature MET141, in the postmenopausal and systemically untreated patients with ER+, HER2 cancers of the METABRIC cohort, where nearly all were node negative. P values are from the Cox proportional hazards model. Continuous risk scores were divided by the standard deviation to directly compare HRs between scores with differently distributed values, and the Cox analysis is limited to 10 years. Results are shown for endpoint breast cancer death. B, Cumulative incidence of breast cancer death in the postmenopausal patients of the METABRIC cohort with ER+, HER2 cancers who did not receive systemic therapy, for each of the 15 previously published gene signatures, AGR, and MET141.

Close modal

Herein, we present transcriptomic analyses of the SweBCG91-RT trial, a trial of early-stage breast cancer with long-term follow-up. As the majority of patients were systemically untreated, this cohort is uniquely suited to address the question of which patients may be spared endocrine therapy. We used comprehensive transcriptomic profiling to evaluate the prognostic performance of 15 previously described breast cancer signatures and we show that although most signatures performed well on the group level, there was considerable discordance on the individual patient level. To overcome this limitation of discordance between individual signatures, we developed the concept of AGR and an associated novel 141-gene signature (MET141), which were independently validated in SweBCG91-RT and in the METABRIC data set. Both AGR and MET141 can identify postmenopausal and systemically untreated patients with ER+, HER2 cancers with excellent prognosis and who may be candidates for omission of systemic therapy, including endocrine therapy. Furthermore, unlike AGR, which requires calculation and summation of risk from 15 different signatures, the MET141 signature distills similar information into a single signature.

The recent EBCTCG meta-analysis showed that late recurrences are a significant clinical problem, and that efforts to avoid endocrine therapy must rely on long-term follow-up data (13). In this study, we show that the performance of calculated scores from previously published signatures deteriorates with longer follow-up. Despite this, many of the signatures can identify a large proportion of patients where over 90% are free of metastasis at 15 years, and rates of death from breast cancer less than 5% at 15 years, even without any systemic therapy. Signatures for treatment prediction are often validated by performing analysis of treatment effect in subgroups. However, for treatment omission, it has been argued that it may be more appropriate to consider the relative treatment effect constant over subgroups and to assess baseline risk (9). This should apply for adjuvant endocrine therapy for breast cancer, where few studies find subgroups within ER+ tumors without any treatment effect, and where a long-term excellent prognosis means modest absolute effect of therapy. Therefore, we here present the long-term results for a low-risk patient subgroup that were not given any systemic adjuvant treatment, and we stratify the results for score quartiles for each signature to allow an unbiased evaluation of what can potentially be achieved by transcriptomic profiling. We deliberately do not select a specific cutoff, as there is no consensus for which rate of metastasis is acceptable, but highly individual and dependent on patient preferences, comorbidities and side effects experienced. However, we chose to highlight results for the lowest risk quartile, where several signatures can identify a group of patients without any metastasis during the first 10 years, and deaths from breast cancer below 5% at 15 years. We believe that these predicted rates may be low enough to discuss omission of endocrine therapy in select patients, but the decision will ultimately be up to the patient and treating physician following a balanced discussion of risks and benefits. Ideally, deescalation of endocrine therapy should be investigated in prospective trials.

An emerging dilemma is the considerable discordance between results of multiple gene-expression tests currently in clinical use and risk prediction for individual patients. Indeed, we have largely confirmed the results by Bartlett and colleagues, in a different cohort, which showed only 39% of patients classified uniformly by five tests as low-risk, whereas individual tests predicted a much larger proportion as low risk. The same authors showed that three different subtyping tests disagreed for 41% of tumors (12). In our current work, we present a strategy of overcoming this by using a whole-transcriptome platform and the average of all the signatures. This approach produced results consistent with the best individual signatures and could potentially improve intersignature variability because it relies on more data points. However, we have included all the signatures in the calculation of AGR and there are likely additional methods or modifications that could further improve risk stratification, such as removal of the signatures with the lowest individual performance or reweighting the signatures. These approaches will be tested in future studies.

Although these data suggest it may be valuable to profile tumors with all available signatures, this is not feasible for numerous practical reasons, including cost and availability of enough sample material from the tumor. To that end, we developed a novel 141-gene signature in publicly available cohorts that is based on genes correlated with AGR. We show that MET141 captures the same biology as the AGR and has a similar performance but would be considerably more feasible in the clinical setting. Although promising in this validation study, it remains to be tested in further patient cohorts if the performance and robustness is superior to currently available signatures.

There are several strengths of this study. First, we utilize a CLIA-certified comprehensive whole-transcriptome approach that produces quality results for FFPE tissue and allows us to assess multiple previously described signatures simultaneously. Further, this study examined a large patient cohort from a well-defined randomized trial. In addition to the benefits of using sample material from an unconfounded randomized phase III trial, the fact that so many of these patients were systemically untreated and followed for such a long time is unique and allows for the novel findings reported herein.

Despite these strengths, there remain some limitations to this study. One limitation is the use of surrogate scores for the previously published signatures. This may produce slightly different scores than using the approved and commercially available diagnostic tests. However, the surrogate scores show the expected high correlation with Ki67 and histologic grade, and we demonstrate that these surrogate scores are able to prognosticate for recurrence risk in two separate data sets, which supports that the calculated scores are incorporating similar information to the clinically used scores. Further, we are not using thresholds originally specified for the individual signatures and the exact definition of low-risk or high-risk tumor may be slightly different in this study. Instead, we group scores by quartiles and when presenting HRs, normalize the scores to the standard deviation of each score. This is done deliberately to directly compare between signatures. If transferring these results to a clinical setting, further stratification by cutoff point determination may be desirable to select those patients at lowest risk for systemic recurrence. Another limitation, inherent in all trials with such long follow-up, is the use of outdated or less relevant treatments as compared with contemporary practice. In this study, however, because we are specifically investigating systemically untreated patients in the adjuvant setting, this is not a major concern. The length of follow-up should not influence the time to metastasis or breast cancer death, except for possible current therapies for treatment of relapses, which could slightly improve the outcome. With regard to radiotherapy, the patients in the trial were randomized to receive either whole breast radiotherapy or no radiotherapy. We chose to combine the RT+ and RT patients in this study to increase power, because the original study did not find difference between ± RT with respect to distant metastasis or death from breast cancer. Besides treatment, baseline risk may change over time due to systematic changes in detection or staging. In the original study, 65% of patients had screen detected tumors and lymph node status was defined based on axillary lymph node dissection, which is likely less sensitive to small-volume lymph node metastases compared with sentinel lymph node biopsies, which is performed today. Thus, if any change in baseline risk, we would anticipate the baseline risk to be even lower.

In conclusion, calculated scores from previously developed breast cancer signatures are largely prognostic in a breast cancer cohort who are node-negative, postmenopausal and systemically untreated with ER+, HER2 tumors. However, the signatures are discordant on an individual patient level, and we therefore propose that an average of the signatures can result in more robust patient-level results. Using this average, or an associated 141-gene signature, patients can be identified with an excellent long-term freedom from metastasis even in the absence of endocrine treatment.

S.L. Chang is an employee/paid consultant for and reports receiving commercial research grants from PFS Genomics, and holds ownership interest (including patents) in PFS Genomics and Decipher Biosciences. N. Fishbane is an employee/paid consultant for and holds ownership interest (including patents) in Decipher Biosciences. E. Davicioni is an employee/paid consultant for Decipher Biosciences. E. Holmberg reports receiving other commercial research support from PFS Genomics research contract. F.Y. Feng holds ownership interest (including patents) in and is an advisory board member/unpaid consultant for PFS Genomics. C.W. Speers holds ownership interest (including patents) in and is an advisory board member/unpaid consultant for PFS Genomics. L.J. Pierce holds ownership interest (including patents) in and is an advisory board member/unpaid consultant for PFS Genomics. P. Malmström reports receiving other remuneration from PFS Genomics (royalty agreement). P. Karlsson reports receiving other remuneration from PFS Genomics (research contract). No potential conflicts of interest were disclosed by the other authors.

Conception and design: M. Sjöström, S.L. Chang, F.Y. Feng, C.W. Speers, L.J. Pierce, P. Malmström, M. Fernö, P. Karlsson

Development of methodology: S.L. Chang, N. Fishbane, E. Davicioni, F.Y. Feng, C.W. Speers

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): E. Davicioni, P. Malmström, M. Fernö, P. Karlsson

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): M. Sjöström, S.L. Chang, N. Fishbane, E. Davicioni, L. Hartman, F.Y. Feng, C.W. Speers, L.J. Pierce, P. Malmström, P. Karlsson

Writing, review, and/or revision of the manuscript: M. Sjöström, S.L. Chang, N. Fishbane, E. Davicioni, L. Hartman, E. Holmberg, F.Y. Feng, C.W. Speers, L.J. Pierce, P. Malmström, M. Fernö, P. Karlsson

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): S.L. Chang, E. Davicioni, L. Hartman, E. Holmberg, F.Y. Feng, M. Fernö, P. Karlsson

Study supervision: F.Y. Feng, P. Malmström, M. Fernö, P. Karlsson

The authors thank Kristina Lövgren for expert technical assistance, Sara Baker for database management and administrative support, and Fredrika Killander for updating the SweBCG91-RT clinical information. This work was supported by PFS Genomics, Swedish Breast Cancer Association (BRO), Swedish Cancer Society, Faculty of Medicine at Lund University, Lund University Research Foundation, Gunnar Nilsson Cancer Foundation, Anna and Edwin Berger Foundation, Swedish Cancer and Allergy Foundation, Skåne County Research Foundation (FOU and PhD studies grant), Mrs. Berta Kamprad Research Foundation, King Gustav V Jubilee Clinic Cancer Foundation in Gothenburg, and the LUA/ALF-agreement in West and South Sweden.

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.

1.
Hosseini
A
,
Khoury
AL
,
Esserman
LJ
. 
Precision surgery and avoiding over-treatment
.
Eur J Surg Oncol
2017
;
43
:
938
43
.
2.
Curigliano
G
,
Burstein
HJ
,
E
PW
,
Gnant
M
,
Dubsky
P
,
Loibl
S
, et al
De-escalating and escalating treatments for early-stage breast cancer: the St. Gallen International Expert Consensus Conference on the Primary Therapy of Early Breast Cancer 2017
.
Ann Oncol
2017
;
28
:
1700
12
.
3.
Gnant
M
,
Sestak
I
,
Filipits
M
,
Dowsett
M
,
Balic
M
,
Lopez-Knowles
E
, et al
Identifying clinically relevant prognostic subgroups of postmenopausal women with node-positive hormone receptor-positive early-stage breast cancer treated with endocrine therapy: a combined analysis of ABCSG-8 and ATAC using the PAM50 risk of recurrence score and intrinsic subtype
.
Ann Oncol
2015
;
26
:
1685
91
.
4.
Sparano
JA
,
Gray
RJ
,
Makower
DF
,
Pritchard
KI
,
Albain
KS
,
Hayes
DF
, et al
Adjuvant chemotherapy guided by a 21-gene expression assay in breast cancer
.
N Engl J Med
2018
;
379
:
111
21
.
5.
Cardoso
F
,
van't Veer
LJ
,
Bogaerts
J
,
Slaets
L
,
Viale
G
,
Delaloge
S
, et al
70-gene signature as an aid to treatment decisions in early-stage breast cancer
.
N Engl J Med
2016
;
375
:
717
29
.
6.
Davies
C
,
Godwin
J
,
Gray
R
,
Clarke
M
,
Cutter
D
,
Darby
S
, et al
Relevance of breast cancer hormone receptors and other factors to the efficacy of adjuvant tamoxifen: patient-level meta-analysis of randomised trials
.
Lancet
2011
;
378
:
771
84
.
7.
Early Breast Cancer Trialists' Collaborative Group (EBCTCG). 
Aromatase inhibitors versus tamoxifen in early breast cancer: patient-level meta-analysis of the randomised trials
.
Lancet
2015
;
386
:
1341
52
.
8.
Chlebowski
RT
,
Geller
ML
. 
Adherence to endocrine therapy for breast cancer
.
Oncology
2006
;
71
:
1
9
.
9.
Peto
R
. 
Current misconception 3: that subgroup-specific trial mortality results often provide a good basis for individualising patient care
.
Br J Cancer
2011
;
104
:
1057
8
.
10.
Ohnstad
HO
,
Borgen
E
,
Falk
RS
,
Lien
TG
,
Aaserud
M
,
Sveli
MAT
, et al
Prognostic value of PAM50 and risk of recurrence score in patients with early-stage breast cancer with long-term follow-up
.
Breast Cancer Res
2017
;
19
:
120
.
11.
Esserman
LJ
,
Yau
C
,
Thompson
CK
,
van 't Veer
LJ
,
Borowsky
AD
,
Hoadley
KA
, et al
Use of molecular tools to identify patients with indolent breast cancers with ultralow risk over 2 decades
.
JAMA Oncology
2017
;
3
:
1503
10
.
12.
Bartlett
JM
,
Bayani
J
,
Marshall
A
,
Dunn
JA
,
Campbell
A
,
Cunningham
C
, et al
Comparing breast cancer multiparameter tests in the OPTIMA prelim trial: no test is more equal than the others
.
J Natl Cancer Inst
2016
;
108
.
13.
Pan
H
,
Gray
R
,
Braybrooke
J
,
Davies
C
,
Taylor
C
,
McGale
P
, et al
20-year risks of breast-cancer recurrence after stopping endocrine therapy at 5 years
.
N Engl J Med
2017
;
377
:
1836
46
.
14.
Sjostrom
M
,
Lundstedt
D
,
Hartman
L
,
Holmberg
E
,
Killander
F
,
Kovacs
A
, et al
Response to radiotherapy after breast-conserving surgery in different breast cancer subtypes in the swedish breast cancer group 91 radiotherapy randomized clinical trial
.
J Clin Oncol
2017
;
35
:
3222
9
.
15.
Killander
F
,
Karlsson
P
,
Anderson
H
,
Mattsson
J
,
Holmberg
E
,
Lundstedt
D
, et al
No breast cancer subgroup can be spared postoperative radiotherapy after breast-conserving surgery. Fifteen-year results from the Swedish Breast Cancer Group randomised trial, SweBCG 91 RT
.
Eur J Cancer
2016
;
67
:
57
65
.
16.
Malmström
P
,
Holmberg
L
,
Anderson
H
,
Mattsson
J
,
Jonsson
PE
,
Tennvall-Nittby
L
, et al
Breast conservation surgery, with and without radiotherapy, in women with lymph node-negative breast cancer: a randomised clinical trial in a population with access to public mammography screening
.
Eur J Cancer
2003
;
39
:
1690
7
.
17.
Gourgou-Bourgade
S
,
Cameron
D
,
Poortmans
P
,
Asselain
B
,
Azria
D
,
Cardoso
F
, et al
Guidelines for time-to-event end point definitions in breast cancer trials: results of the DATECAN initiative (definition for the assessment of time-to-event endpoints in CANcer trials)
.
Ann Oncol
2015
;
26
:
873
9
.
18.
Hudis
CA
,
Barlow
WE
,
Costantino
JP
,
Gray
RJ
,
Pritchard
KI
,
Chapman
JA
, et al
Proposal for standardized definitions for efficacy end points in adjuvant breast cancer trials: the STEEP system
.
J Clin Oncol
2007
;
25
:
2127
32
.
19.
Piccolo
SR
,
Sun
Y
,
Campbell
JD
,
Lenburg
ME
,
Bild
AH
,
Johnson
WE
. 
A single-sample microarray normalization method to facilitate personalized-medicine workflows
.
Genomics
2012
;
100
:
337
44
.
20.
Curtis
C
,
Shah
SP
,
Chin
SF
,
Turashvili
G
,
Rueda
OM
,
Dunning
MJ
, et al
The genomic and transcriptomic architecture of 2,000 breast tumours reveals novel subgroups
.
Nature
2012
;
486
:
346
52
.
21.
Paik
S
,
Shak
S
,
Tang
G
,
Kim
C
,
Baker
J
,
Cronin
M
, et al
A multigene assay to predict recurrence of tamoxifen-treated, node-negative breast cancer
.
N Engl J Med
2004
;
351
:
2817
26
.
22.
Filipits
M
,
Rudas
M
,
Jakesz
R
,
Dubsky
P
,
Fitzal
F
,
Singer
CF
, et al
A new molecular predictor of distant recurrence in ER-positive, HER2-negative breast cancer adds independent information to conventional clinical risk factors
.
Clin Cancer Res
2011
;
17
:
6012
20
.
23.
Sotiriou
C
,
Wirapati
P
,
Loi
S
,
Harris
A
,
Fox
S
,
Smeds
J
, et al
Gene expression profiling in breast cancer: understanding the molecular basis of histologic grade to improve prognosis
.
J Natl Cancer Inst
2006
;
98
:
262
72
.
24.
Parker
JS
,
Mullins
M
,
Cheang
MC
,
Leung
S
,
Voduc
D
,
Vickery
T
, et al
Supervised risk predictor of breast cancer based on intrinsic subtypes
.
J Clin Oncol
2009
;
27
:
1160
7
.
25.
van de Vijver
MJ
,
He
YD
,
van't Veer
LJ
,
Dai
H
,
Hart
AA
,
Voskuil
DW
, et al
A gene-expression signature as a predictor of survival in breast cancer
.
N Engl J Med
2002
;
347
:
1999
2009
.
26.
Haibe-Kains
B
,
Desmedt
C
,
Rothe
F
,
Piccart
M
,
Sotiriou
C
,
Bontempi
G
. 
A fuzzy gene expression-based computational approach improves breast cancer prognostication
.
Genome Biol
2010
;
11
:
R18
.
27.
Loi
S
,
Haibe-Kains
B
,
Desmedt
C
,
Wirapati
P
,
Lallemand
F
,
Tutt
AM
, et al
Predicting prognosis using molecular profiling in estrogen receptor-positive breast cancer treated with tamoxifen
.
BMC Genomics
2008
;
9
:
239
.
28.
Wang
Y
,
Klijn
JG
,
Zhang
Y
,
Sieuwerts
AM
,
Look
MP
,
Yang
F
, et al
Gene-expression profiles to predict distant metastasis of lymph-node-negative primary breast cancer
.
Lancet
2005
;
365
:
671
9
.
29.
Loi
S
,
Haibe-Kains
B
,
Majjaj
S
,
Lallemand
F
,
Durbecq
V
,
Larsimont
D
, et al
PIK3CA mutations associated with gene signature of low mTORC1 signaling and better outcomes in estrogen receptor-positive breast cancer
.
PNAS
2010
;
107
:
10208
13
.
30.
Tutt
A
,
Wang
A
,
Rowland
C
,
Gillett
C
,
Lau
K
,
Chew
K
, et al
Risk estimation of distant metastasis in node-negative, estrogen receptor-positive breast cancer patients using an RT-PCR based prognostic expression signature
.
BMC Cancer
2008
;
8
:
339
.
31.
Davis
LM
,
Harris
C
,
Tang
L
,
Doherty
P
,
Hraber
P
,
Sakai
Y
, et al
Amplification patterns of three genomic regions predict distant recurrence in breast carcinoma
.
J Mol Diagn
2007
;
9
:
327
36
.
32.
Ring
BZ
,
Seitz
RS
,
Beck
R
,
Shasteen
WJ
,
Tarr
SM
,
Cheang
MC
, et al
Novel prognostic immunohistochemical biomarker panel for estrogen receptor-positive breast cancer
.
J Clin Oncol
2006
;
24
:
3039
47
.
33.
Ma
XJ
,
Salunga
R
,
Dahiya
S
,
Wang
W
,
Carney
E
,
Durbecq
V
, et al
A five-gene molecular grade index and HOXB13:IL17BR are complementary prognostic factors in early stage breast cancer
.
Clin Cancer Res
2008
;
14
:
2601
8
.
34.
Bayani
J
,
Yao
CQ
,
Quintayo
MA
,
Yan
F
,
Haider
S
,
D'Costa
A
, et al
Molecular stratification of early breast cancer identifies drug targets to drive stratified medicine
.
NPJ Breast Cancer
2017
;
3
:
3
.
35.
Ma
XJ
,
Wang
Z
,
Ryan
PD
,
Isakoff
SJ
,
Barmettler
A
,
Fuller
A
, et al
A two-gene expression ratio predicts clinical outcome in breast cancer patients treated with tamoxifen
.
Cancer Cell
2004
;
5
:
607
16
.
36.
Gray
B
. 
cmprsk: Subdistribution analysis of competing risks
.
R package version 2.2-7. Available at
: http://CRAN.R-project.org/package=cmprsk.
Accessed May 17
, 
2019
.
37.
Scrucca
L
,
Santucci
A
,
Aversa
F
. 
Competing risk analysis using R: an easy guide for clinicians
.
Bone Marrow Transplant
2007
;
40
:
381
7
.
38.
Schoenfeld
D
. 
Partial residuals for the proportional hazards regression model
.
Biometrika
1982
;
69
:
239
41
.
39.
McShane
LM
,
Altman
DG
,
Sauerbrei
W
,
Taube
SE
,
Gion
M
,
Clark
GM
. 
REporting recommendations for tumour MARKer prognostic studies (REMARK)
.
Eur J Cancer
2005
;
41
:
1690
6
.
40.
Benjamini
Y
,
Hochberg
Y
. 
Controlling the false discovery rate: a practical and powerful approach to multiple testing
.
J R Stat Soc Series B Methodol
1995
;
57
:
289
300
.
41.
Heagerty
PJ
,
Lumley
T
,
Pepe
MS
. 
Time-dependent ROC curves for censored survival data and a diagnostic marker
.
Biometrics
2000
;
56
:
337
44
.
42.
Mi
H
,
Huang
X
,
Muruganujan
A
,
Tang
H
,
Mills
C
,
Kang
D
, et al
PANTHER version 11: expanded annotation data from Gene Ontology and Reactome pathways, and data analysis tool enhancements
.
Nucleic Acids Res
2017
;
45
:
D183
d9
.
43.
Fabregat
A
,
Jupe
S
,
Matthews
L
,
Sidiropoulos
K
,
Gillespie
M
,
Garapati
P
, et al
The reactome pathway knowledgebase
.
Nucleic Acids Res
2018
;
46
:
D649
d55
.
44.
Milacic
M
,
Haw
R
,
Rothfels
K
,
Wu
G
,
Croft
D
,
Hermjakob
H
, et al
Annotating cancer variants and anti-cancer therapeutics in reactome
.
Cancers
2012
;
4
:
1180
211
.
45.
Gendoo
DM
,
Ratanasirigulchai
N
,
Schroder
MS
,
Pare
L
,
Parker
JS
,
Prat
A
, et al
Genefu: an R/Bioconductor package for computation of gene expression-based signatures in breast cancer
.
Bioinformatics
2016
;
32
:
1097
9
.
46.
van 't Veer
LJ
,
Dai
H
,
van de Vijver
MJ
,
He
YD
,
Hart
AA
,
Mao
M
, et al
Gene expression profiling predicts clinical outcome of breast cancer
.
Nature
2002
;
415
:
530
6
.
47.
Servant
N
,
Bollet
MA
,
Halfwerk
H
,
Bleakley
K
,
Kreike
B
,
Jacob
L
, et al
Search for a gene expression signature of breast cancer local recurrence in young women
.
Clin Cancer Res
2012
;
18
:
1704
15
.
48.
Kao
KJ
,
Chang
KM
,
Hsu
HC
,
Huang
AT
. 
Correlation of microarray-based breast cancer molecular subtypes and clinical outcomes: implications for treatment optimization
.
BMC Cancer
2011
;
11
:
143
.
49.
Symmans
WF
,
Hatzis
C
,
Sotiriou
C
,
Andre
F
,
Peintinger
F
,
Regitnig
P
, et al
Genomic index of sensitivity to endocrine therapy for breast cancer
.
J Clin Oncol
2010
;
28
:
4111
9
.