Gene expression profiling to predict the risk of locoregional recurrence ...

Breast Cancer Res Treat (2014) 148:599–613 DOI 10.1007/s10549-014-3188-z

CLINICAL TRIAL

Gene expression profiling to predict the risk of locoregional recurrence in breast cancer: a pooled analysis C. A. Drukker • S. G. Elias • M. V. Nijenhuis • J. Wesseling • H. Bartelink P. Elkhuizen • B. Fowble • P. W. Whitworth • R. R. Patel • F. A. de Snoo • L. J. van ’t Veer • P. D. Beitsch • E. J. Th. Rutgers

•

Received: 28 October 2014 / Accepted: 29 October 2014 / Published online: 21 November 2014 Ó Springer Science+Business Media New York 2014

Abstract The 70-gene signature (MammaPrintTM) has been developed to predict the risk of distant metastases in breast cancer and select those patients who may benefit from adjuvant treatment. Given the strong association between locoregional and distant recurrence, we hypothesize that the 70-gene signature will also be able to predict the risk of locoregional recurrence (LRR). 1,053 breast cancer patients primarily treated with breast-conserving treatment or mastectomy at the Netherlands Cancer Institute between 1984 and 2006 were included. Adjuvant treatment consisted of radiotherapy, chemotherapy, and/or endocrine therapy as indicated by guidelines used at the time. All patients were included in various 70-gene signature validation studies. After a median follow-up of 8.96 years with 87 LRRs, patients with a high-risk 70-gene signature (n = 492) had an

C.A. Drukker, S.G. Elias, and M.V. Nijenhuis have contributed equally to this study. The data in this study have been presented in a poster presentation at the San Antonio Breast Cancer Symposium in December 2012, in an oral presentation at the Annual meeting of the Dutch Society of Surgery of 2012, and the ECCO meeting of 2013. C. A. Drukker M. V. Nijenhuis E. J. Th. Rutgers (&) Department of Surgical Oncology, Netherlands Cancer Institute, Postbus 90203, 1006 BE Amsterdam, The Netherlands e-mail: [email protected]

LRR risk of 12.6 % (95 % CI 9.7–15.8) at 10 years, compared to 6.1 % (95 % CI 4.1–8.5) for low-risk patients (n = 561; P \ 0.001). Adjusting the 70-gene signature in a competing risk model for the clinicopathological factors such as age, tumour size, grade, hormone receptor status, LVI, axillary lymph node involvement, surgical treatment, endocrine treatment, and chemotherapy resulted in a multivariable HR of 1.73 (95 % CI 1.02–2.93; P = 0.042). Adding the signature to the model based on clinicopathological factors improved the discrimination, albeit non-significantly [C-index through 10 years changed from 0.731 (95 % CI 0.682–0.782) to 0.741 (95 % CI 0.693–0.790)]. Calibration of the prognostic models was excellent. The 70-gene signature is an independent prognostic factor for LRR. A significantly lower local recurrence risk was seen in patients with a low-risk 70-gene signature compared to those with high-risk 70-gene signature. Keywords Breast cancer Risk prediction Locoregional recurrence 70-gene signature Surgery Radiation oncology

H. Bartelink P. Elkhuizen Department of Radiation Oncology, Netherlands Cancer Institute, Amsterdam, The Netherlands

M. V. Nijenhuis e-mail: [email protected]

B. Fowble Department of Radiation Oncology, University of California San Francisco, San Francisco, USA

S. G. Elias Julius Center for Health Sciences and Primary Care, University Medical Center Utrecht, Utrecht, The Netherlands

P. W. Whitworth Department of Surgery, Nashville Breast Centre, Nashville, USA

J. Wesseling L. J. van ’t Veer Department of Pathology, Netherlands Cancer Institute, Amsterdam, The Netherlands

R. R. Patel Department of Radiation Oncology, El Camino Hospital, Mountain View, USA

123

600

Introduction For the majority of breast cancer patients, locoregional recurrence (LRR) is becoming a less-frequent clinical problem. Improvements in patient selection, surgical treatment, radiotherapy, and adjuvant systemic therapy have led to a substantial decrease in LRR incidence rates [1, 2]. On average, 3 % of the patients do experience LRR within 5 years after diagnosis [2]. The favourable LRR rates especially raise the issue of overtreatment in women with low LRR risk. Nevertheless, improved identification of high-risk patients is still of importance, not only for local control, but also because local failure is associated with poor survival [3]. Ideally, one would aim to identify patients at a high risk of LRR to better guide optimal locoregional treatment with more extensive surgery, radiotherapy, and adjuvant systemic treatment, while at the same time identification of patients at a low risk of LRR can help to avoid unnecessary adjuvant radiotherapy. Currently, several factors are used to predict LRR and indicate therapy aimed at locoregional control. The most important risk factors for LRR are margin involvement, use of radiotherapy after breast-conserving surgery, and age (\40 years) [4]. Other clinicopathological factors that are used are grade, tumour size, hormonal status, lymphovascular invasion (LVI), and tumour involvement of axillary lymph nodes. Although these factors can be used to predict LRR risk, not all patients can be classified to be at clinically relevant high or low LRR risk. The biological background of breast cancer may help in further assessing the risk of recurrence [5]. Gene expression classifiers, such as the 70-gene signature (MammaPrintTM, Agendia Inc, Amsterdam, the Netherlands), are proven useful additional tools for assessing distant recurrence risk in breast cancer [6–8]. The 70-gene signature has been extensively validated on retrospective as well as prospective data [7–10]. The results show a favourable 5-year distant recurrence-free-interval for patients with a low-risk 70-gene signature, even in the absence of adjuvant systemic treatment and despite poor clinicopathological factors [7]. Since LRR is an independent predictor of subsequent distant metastases [11], we hypothesize that the 70-gene signature will be able to predict the risk of LRR as well. F. A. de Snoo L. J. van ’t Veer Agendia Inc, Amsterdam, The Netherlands L. J. van ’t Veer Department of Laboratory Medicine, University of California, San Francisco, USA P. D. Beitsch Dallas Surgical Group, Dallas Breast Centre, Dallas, USA

123

Breast Cancer Res Treat (2014) 148:599–613

The aim of this study is to evaluate the performance of the 70-gene signature in the prediction of LRR and to evaluate the additional prognostic value of the 70-gene signature to clinicopathological factors that are currently used.

Methods All 1,053 individual breast cancer patients part of eight previous 70-gene signature validation studies, but not included in its development, and treated at the Netherlands Cancer Institute (NKI) were eligible for the current study (Study flow chart, Fig. 1). Only NKI-treated patients were included to facilitate standardized ascertainment of locoregional events during the extended follow-up period and to allow an update of radiotherapy information. All patients had tumour-free resection margins. Details of study design, rationale, and patient eligibility of seven of the included studies have been described elsewhere [8, 9, 12–16]. In short, all patients were women with histologically proven, operable, invasive breast cancer (T1-3N0-1M0), diagnosed between 1984 and 2006. All patients were primarily treated with mastectomy or breast-conserving therapy (BCT), comprising lumpectomy followed by radiotherapy. Radiotherapy consisted mostly of 50 Gy, 25 fractions wholebreast irradiation, with or without a 16 Gy-boost dose after BCT or chest wall irradiation after mastectomy and/or internal mammary chain radiotherapy. Adjuvant systemic treatment consisted of chemotherapy and/or endocrine therapy as indicated by guidelines used at the time. The 70-gene signature did not influence treatment decisions except to some extent in the single prospective study [12]. One of the eight studies was not yet published at the time of these analyses were performed. In this study, Bedard et al. included 320 women aged C65 years and diagnosed with early-stage breast cancer (T1-3N0-1M0) between 1987 and 2003 at the NKI or Jules Bordet Institute (Brussels, Belgium). None of these patients received adjuvant chemotherapy. All individual studies complied with Ethical Review Board standards. Clinicopathological factors and the 70-gene signature Information on age, grade, oestrogen receptor (ER), Human Epidermal Growth Factor Receptor 2 (HER2), LVI, tumour size, and number of involved axillary lymph nodes was derived from the original study data that included pathological review by an expert (Hans Peterse). LVI was not documented in all studies; therefore, an experienced breast pathologist (JW) reviewed 150 of the missing cases for this study. Frozen tumour samples from each patient were processed at the Agendia laboratory (Amsterdam, the Netherlands) for RNA isolation, amplification, and


601

Fig. 1 Study flow chart

labelling as described elsewhere [6, 8, 17]. To assess the mRNA expression level of the 70-genes, RNA was hybridized to a custom-designed array, commercially available as MammaPrintTM. Agendia Inc. is ISO17025 certified, CLIA accredited, and FDA cleared. Tumours were classified as 70-gene signature low (indexscore [ 0.4) or high (index-score B 0.4) risk at the time of the initial studies [6, 17]. Follow-up Follow-up for locoregional recurrence and death was updated through November 2011 using data from the NKI Tumour Registry complemented with review of the original patient records. Locoregional recurrence was defined as reappearance of breast cancer in the ipsilateral breast or chest wall or ipsilateral regional lymph node involvement, six months or longer after diagnosis. Ipsilateral supraclavicular lymph node involvement was included as regional recurrence throughout the follow-up period, although TNM-editions 4 and 5 considered such recurrences M1 instead of N3. Statistical analysis Appendix 1 provides the detailed statistical methods. In short, competing risk analyses were used as not to overestimate the absolute LRR risk [18]. Follow-up started at diagnosis and ended at the first manifestation of LRR (event) or death (competing event), or at the end of the

follow-up without LRR or death (censored). Occurrences of distant metastasis, contralateral breast cancer, or second primary tumours were not considered censoring events or competing risks. The univariable 5- and 10-year absolute risk of LRR for the 70-gene signature high- and low-risk groups was estimated using the cumulative incidence function [19] and compared using Gray’s test [20]. Multivariable analyses were performed using Fine and Gray competing risk regression [21]. A multivariable model was constructed comprising solely of routine clinicopathological factors and treatment. The 70-gene signature was added to this model to evaluate its additional and independent prognostic value. The combined prognostic performance of the multivariable models was evaluated for discrimination (Harrell’s C-index adapted to competing risk analyses [18] ) and calibration. Model improvement upon the addition of the 70-gene signature was tested by the pseudo-likelihood ratio test, by change in C-index and by net reclassification improvement (NRI) measures adapted to time-to-event data with censoring [22]. Proportionality of the hazards was tested for the 70-gene signature by a time-covariate interaction. Multiplicative interaction terms were also evaluated for the 70-gene signature with age and with primary locoregional treatment. Analyses were performed using R version 3.0.1. Missing data were multiply imputed. All statistical tests were two-sided with a cut-off for statistical significance of 5 %, without accounting for multiple testing. For categorical NRI measures, low risk was defined as a risk of LRR \2.5 % at 5 years, because all patients received breast irradiation after breast-conserving surgery which at

123

602

least halves the recurrence risk. An annual LRR risk of 1 % is generally accepted without the need for radiotherapy [3, 23]. High risk was defined as an LRR risk of [7.5 % at 5 years, because these patients will not only benefit from adjuvant treatment in achieving locoregional control, but also to optimize overall survival [24].

Results Patient and tumour characteristics by 70-gene signature Of the 1,053 patients with primary operable breast cancer included in this pooled study, 561 (53 %) were 70-gene signature low risk and 492 (47 %) were high risk. Low-risk patients were generally older at the time of diagnosis, had smaller tumours with lower grade, which were more often oestrogen receptor (ER) and progesterone receptor (PR) positive, HER2 negative, and without lymphovascular invasion as compared to the high-risk patients (Table 1). In line with the more aggressive phenotype, more chemotherapy was administered to 70-gene signature high-risk patients, while no difference was seen for endocrine therapy. Neither regarding their definitive surgical treatment nor regarding the administration of local therapy (overall and according to surgery) was a difference seen between high- and lowrisk patients. Absolute risk of LRR, overall and according to the 70gene signature Through 10 years of follow-up (median 8.96 years; IQR 5.3–10), 87 patients developed LRR and 204 patients died without LRR. The overall LRR risk was 5.7 % (95 % CI 4.4–7.2) at 5 years and 9.1 % (95 % CI 7.4–11.1) at 10 years (Fig. 4). First locoregional relapse was local in 53, regional in 4, and synchronous locoregional in 30 patients. Distant metastases were diagnosed in 246 patients, including 57 of LRR cases, of which 17 occurred before LRR (preceding LRR by a median of 1 month). The 70-gene signature was strongly associated with LRR risk without taking other prognostic factors into account (Fig. 2a). In the entire cohort, the 5-year LRR risk was 2.7 % (95 % CI 1.6–4.3) for 70-gene signature lowrisk patients and 9.1 % (95 % CI 6.8–11.9) for high-risk patients, which was 6.1 % (95 % CI 4.1–8.5) and 12.6 % (95 % CI 9.7–15.8) at 10 years, respectively (P \ 0.001; Fig. 2a). Results were similar when stratified by surgery and radiotherapy, albeit less pronounced and non-

123


significant in the mastectomy group without radiotherapy (Fig. 2b–d). Additional value of the 70-gene signature to standard clinicopathological factors The results of the competing risk regression analyses are shown in Table 2. Univariably, patients with a high-risk 70-gene signature had a 2.40 times (95 % CI 1.54–3.74) higher risk of LRR than patients with a low-risk 70-gene signature (P \ 0.001). Adjusting the 70-gene signature for the standard clinicopathological factors led to a multivariable HR of 1.73 (95 % CI 1.02–2.93; P = 0.042). This HR was not different in (predefined) subgroups of age and primary locoregional treatment (interaction with age (C50 vs. \50): P = 0.89; interaction with primary locoregional treatment (breast conserving versus mastectomy with or without local radiotherapy): P = 0.13); however, the effect of the 70-gene signature did depend on the time since diagnosis, with a strong effect in the first 5 years (HR 2.59; 95 % CI 1.39–4.83; P = 0.003), and no relation at longer follow-up (interaction with time: P = 0.019). Below results are therefore focussed on the results taking this 70-gene signature time dependency into account. Other significant independent prognostic factors for LRR were LVI (HR 1.88; 95 % CI 1.12–3.15; P = 0.016), administration of chemotherapy (HR 0.50; 95 % CI 0.26–0.98; P = 0.035), and age (P = 0.001). With regard to age, the adjusted LRR risk increased with younger age below the age of 40, remained stable between ages 40 to 50, and decreased with age above 50 towards a negligible risk beyond 70 years of age (HRs compared to age 50 approaching zero; Figs. 3, 4). The models with and without the 70-gene signature showed good agreement between predicted and observed LRR risk at 5 and 10 years, respectively (see Fig. 5 for calibration plots). Based on the C-index, the discrimination for LRR within five years post-diagnosis increased from 0.760 (95 % CI 0.700–0.820) to 0.776 (95 % CI 0.723–0.829) when adding the 70-gene signature to the model with clinicopathological factors, at 10-years, the Cindex increased from 0.731 (95 % CI 0.682–0.782) to 0.744 (95 % CI 0.696–0.791; Table 2). These improvements in discrimination were however not significant (P = 0.30 and 0.54 respectively). Evaluating the multivariable models as developed in the entire cohort for discrimination in patient subgroups showed that the C-index also improved by extending the model with clinicopathological factors with the 70-gene signature in patients treated with BCT and in patients treated with mastectomy and local radiotherapy, but not in patients following


603

Table 1 Patient and tumour characteristics of 1,053 primary operable breast cancer patients treated at the Netherlands Cancer Institute stratified by 70-gene signature high and low risk 70-GS low risk N = 561 (53 %)

70-GS high risk N = 492 (47 %)

Table 1 continued 70-GS low risk N = 561 (53 %)

P

[50 years

56 (27–96)

51 (26–88)

\0.001

Radiotherapy in breast-conserving group

188 (34 %)

232 (47 %)

\0.001

Radiotherapy in mastectomy group

373 (66 %)

260 (53 %)

1

0

0

22 (12)

24 (11)

1

0

Grade 1

233 (43 %)

52 (11 %)

Grade 2

259 (48 %)

154 (32 %)

Grade 3

48 (9 %)

275 (57 %)

21

11

No

305 (73 %)

272 (65 %)

Yes

111 (27 %)

145 (35 %)

145

75

249 (98 %)

216 (100 %)

0.38

151 (51 %)

147 (57 %)

0.13

10

18

No

298 (53 %)

280 (57 %)

Yes

263 (47 %)

212 (43 %)

0

0

Adjuvant endocrine therapy

Tumour characteristics Tumour size in mm, mean (SD)

P

Radiotherapy according to surgery

Patient characteristics Age at diagnosis in years, median (min–max) B50 years

70-GS high risk N = 492 (47 %)

0.002

Adjuvant chemotherapy No

Tumour grade \0.001

Yes

456 (81 %)

304 (62 %)

105 (19 %)

188 (38 %)

0

0

0.24

\0.001

70-GS 70-gene signature, mm millimetre, SD standard deviation a

Italic numbers denote absolute number of cases with missing values for the respective variable

Lymphovascular invasion 0.013

b

P-values are based on the Student’s t Test (normally distributed continuous data), the Mann–Whitney test (non-normally distributed continuous data), or the Fisher’s Exact test (categorical data)

Axillary status Node negative

286 (51 %)

224 (46 %)

Node positive

271 (49 %)

266 (54 %)

4

2

0.073

Oestrogen receptor status Negative

16 (3 %)

173 (35 %)

Positive

542 (97 %)

318 (65 %)

3

1

\0.001

Progesterone receptor status Negative

97 (18 %)

255 (53 %)

Positive

448 (82 %) 16

223 (47 %) 14

Negative

441 (96 %)

320 (76 %)

Positive

20 (4 %)

100 (24 %)

100

72

Breast conserving

257 (46 %)

224 (46 %)

Mastectomy

302 (54 %)

265 (54 %)

2

3

No

151 (27 %)

112 (24 %)

Yes

246 (45 %)

232 (49 %)

\0.001

HER2 status \0.001

Treatment Surgical procedure 1.00

Radiotherapy

Yes, with boost

153 (28 %)

129 (27 %)

11

19

0.29

mastectomy without radiotherapy. The improvement in discrimination was statistically significant in the mastectomy group with radiotherapy at 5 years (Table 3). With regard to LRR risk reclassification in the entire cohort at 5 years, 23 % of patients had a low predicted LRR risk (\2.5 %) according to the model based on clinicopathological factors, 54 % an intermediate LRR risk, and 23 % a high LRR risk (C7.5 %) (Table 4a). Adding the 70-gene signature to that model resulted in a reclassification of 30 % of all patients, especially downward (19 %). Based on this extended model, 31 % of patients had a low and 24 % a high predicted 5-year LRR risk. Reclassification was largest in the intermediate risk group according to clinicopathological factors only (37 %, with 24 % moving down; Table 4a). In absolute terms, the largest amount of reclassification in this intermediate risk group occurred in patients expected to remain free from LRR, of whom a net 13 % (95 % CI 6–19 %; P \ 0.001) were correctly reclassified downwards (the NRI in those expected to experience LRR was 16 % (95 % CI 11–44 %; P = 0.25), leading to an overall NRI clinical of 29 % (95 % CI 1–57 %; P = 0.047); Table 4b). Of the patients moving down in this intermediate risk group, 75 % had received radiotherapy (Table 4a). Evaluation of reclassification in the BCT group, whom all had received

123

604


Fig. 2 Cumulative risk of locoregional recurrence through 10 years of follow-up for the entire cohort of 1,053 primary operable breast cancer patients treated at the Netherlands Cancer Institute (a), after breast-conserving therapy (b), and after mastectomy without (c) or

with radiotherapy (d), stratified for 70-gene signature low risk and high risk. Abbreviations: 70-gene signature (70-GS), breast-conserving therapy (BCT), confidence interval (CI), radiotherapy (RT)

radiotherapy by default, showed that 59 % had an intermediate 5-year LRR risk according to clinicopathological factors, of whom 22 % were reclassified downward by adding the 70-gene signature (Table 5a). Here, the net improvement for those expected to remain free from LRR was 11 % (95 % CI 2–20; P = 0.02; Table 5B). Net reclassification improvement in patients treated by mastectomy with or without radiotherapy was similar to what was observed for the entire cohort, especially regarding reclassification of patients expected to remain free from LRR (Fig. 6a, c, e). Furthermore, and congruent with the lack of association with LRR risk from 5 through 10 years after diagnosis, reclassification performance of

the 70-gene signature at 10 years was lower than at 5 years (Fig. 6b, d, f).

123

Discussion In this pooled analysis of eight studies, the 70-gene signature showed to be the strongest independent prognostic factor for LRR in the first five years following adequate primary breast cancer treatment with BCT or mastectomy. From previous studies, it is known that LRR rates are highest in the first 5 years following diagnosis, and 59 of 87 (68 %) of the observed LRRs in our study occurred


605

Table 2 Risk of locoregional recurrence within 10 years of breast cancer diagnosis according to the 70-gene signature and routine clinicopathological variables—Fine and Gray competing risk Univariable analysis

regression analysis of pooled data from eight individual studies, totalling 1,053 primary operable breast cancer patients treated at the Netherlands Cancer Institute

Multivariable analyses Model based on clinicopathological factors

70-gene signature extended models No time dependency

Time dependent

HR (95 % CI)

HR (95 % CI)

P

HR (95 % CI)

P

1.73 (1.02–2.93)

0.042

HR (95 % CI)

P

P

2.40 (1.54–3.74)

\0.001

High vs low risk, \ 5 years

3.49 (1.95–6.26)

\0.001

2.59 (1.39–4.83)

0.003

High vs low risk, 5-10 years

1.22 (0.58–2.54)

0.60

0.84 (0.37–1.92)

0.68

Tumour size, per 5 mm

1.04 (0.97–1.12)

0.24

1.03 (0.93–1.14)

0.60

1.03 (0.93–1.15)

0.56

1.03 (0.93–1.15)

0.57

Grade 2 vs 1

1.76 (0.93–3.33) 2.91 (1.56–5.43)

0.082

1.57 (0.82–3.00) 2.12 (1.01–4.45)

0.17

1.41 (0.72–2.78) 1.60 (0.72–3.56)

0.32

1.42 (0.72–2.80) 1.61 (0.72–3.57)

0.31

70-gene signature High vs low risk (through 10 years) Time-dependent, follow-up time

Tumour characteristics

Grade 3 vs 1

\0.001

0.047

0.25

0.25

Lymphovascular invasion, yes vs no

1.83 (1.17–2.85)

0.008

1.94 (1.16–3.25)

0.012

1.87 (1.12–3.15)

0.018

1.88 (1.12–3.15)

0.016

Axillary status, per positive node

1.04 (0.94–1.14) 0.55 (0.34–0.88)

0.47

1.07 (0.95–1.21) 0.86 (0.47–1.58)

0.28

1.07 (0.95–1.21) 0.96 (0.53–1.74)

0.26

1.07 (0.95–1.21) 0.97 (0.54–1.75)

0.26

Oestrogen receptor, pos. vs neg.

0.014

0.63

0.89

0.92

Treatment Mastectomy vs breast conserving

1.07 (0.70–1.63)

0.75

0.72 (0.39–1.33)

0.29

0.72 (0.39–1.34)

0.31

0.72 (0.39–1.34)

0.31

Radiotherapy, yes vs no

0.76 (0.48–1.21)

0.25

0.69 (0.32–1.48)

0.34

0.68 (0.32–1.44)

0.31

0.68 (0.32–1.43)

0.30

Radiotherapy boost, yes vs no

0.83 (0.50–1.36)

0.46

0.84 (0.47–1.49)

0.54

0.86 (0.48–1.52)

0.60

0.86 (0.49–1.53)

0.61

Endocrine therapy, yes vs no

0.51 (0.32–0.80)

0.004

0.67 (0.40–1.11)

0.12

0.67 (0.40–1.12)

0.12

0.66 (0.40–1.10)

0.11

Chemotherapy, yes vs no

1.09 (0.69–1.72)

0.71

0.53 (0.27–1.02)

0.058

0.51 (0.26–0.98)

0.042

0.50 (0.26–0.95)

0.035

–

0.001

–

0.001

Age, restricted cubic spline1

–

\0.001

\0.001

–

Model comparison and performance Pseudo-likelihood ratio test vs traditional predictor model

–

Ref.

P = 0.042

P = 0.008

vs 70-GS not time-dependent model

–

–

Ref.

P = 0.019

Discrimination C-index (95 % CI) through 5 years

–

0.760 (0.700–0.820)

0.773 (0.717–0.830)

0.776 (0.723–0.829)

C-index (95 % CI) through 10 years

–

0.731 (0.680–0.782)

0.741 (0.693–0.790)

0.744 (0.696–0.791)

70-GS 70-gene signature, CI confidence interval, HR hazard ratio, mm millimetre See Fig. 3 for relation between age and LRR risk

a

123

606

Fig. 3 The relation between age at diagnosis and locoregional recurrence risk (analysed by Fine and Gray competing risk regression). The curve is based on the full cohort of 1,053 primary operable breast cancer cases treated at the Netherlands Cancer Institute. The plotted hazard ratio (95 % confidence interval (CI)) is adjusted for all traditional predictors of LRR with age 50 as reference. Age was modelled with a linear term and the use of a restricted cubic spline function (total 4 degrees of freedom; the spline function significantly improved the model fit (pseudo-likelihood ratio test P = 0.049))

during this time, making LRR risk prediction for this period especially important [3, 24]. The LRR discrimination improved by adding the 70-gene signature to a multivariable model with standard clinicopathological factors, albeit non-significantly. Risk reclassification did however improve significantly upon extension with the 70-gene signature, especially in those at intermediate LRR risk according to the standard clinicopathological factors, and with the largest impact on correct downward reclassification. The results of our study thus support the potential clinical value for the 70-gene signature in LRR prognostication. This pooled study is currently the largest breast cancer patient cohort with long-term LRR follow-up for whom gene expression data are available as well as detailed data on relevant clinicopathological prognostic factors, including therapy. These clinicopathological factors showed the expected direction and magnitude of their relation with LRR risk in our analyses, indicating that all available prognostic information was adequately taken into account and—thus—any shared information with the 70-gene signature. Nevertheless, most established clinicopathological factors did not reach statistical significance due to the limited power caused by the low number of LRR events. Therefore, we used multiple imputation of any missing data

123


Fig. 4 Overall cumulative risk of locoregional recurrence and death as a competing cause through 10 years of follow-up as observed in 1,053 primary operable breast cancer cases treated at the Netherlands Cancer Institute

to maximally preserve sample size and to limit selection bias [25]. Most clinicopathological factors had few missing data except LVI, which still showed to be strongly related with LRR risk in the analyses following multiple imputation. Besides evaluating the improvement of discrimination as measured by the C-index upon extension of a model with standard clinicopathological factors with the 70-gene signature, we assessed net reclassification improvement measures that are more directly interpretable for clinical value. Especially, if they are based on absolute risk categories that inform medical decisions. It should however be noted that external validation of these findings is warranted: First, because the patients included in our pooled dataset were all selected for 70-gene signature validation studies aimed at specific subgroups of patients. This may have resulted in a patient-mix not fully representative of the entire spectrum of breast cancer patients. Then, all patients were treated between 1984 and 2006. During these 22 years, not only adjuvant systemic treatment options improved, but also their indication broadened. As many more contemporary patients receive adjuvant systemic treatment, already lowering their LRR risk, the 70-gene signature may currently have even more impact in identifying women in whom radiotherapy may be


Fig. 5 Calibration curves (i.e. mean predicted vs. observed locoregional recurrence risk) of the clinicopathological model at 5 and 10 years (a and b), of the 70-gene signature extended model at 5 and 10 years (c and d), and of the 70-gene signature extended model taking the time-dependent effect of the signature into account (e and f). Curves are based on the full cohort of 1,053 primary operable breast cancer cases treated at the Netherlands Cancer Institute, using Fine and Gray competing risk regression. Dots depict observed risk

607

within groups of predicted risk (bars are 95 % confidence intervals). Dashed line shows perfect calibration, and black solid line shows inverse variance-weighted linear regression line through data points. Histograms depict distribution of predicted risk among patients with locoregional recurrence (top) or among patients who were censored or who died before any locoregional recurrence (bottom) within the evaluated time period

123

608

Fig. 6 Net reclassification improvement (NRI) of locoregional recurrence (LRR) risk following the time-dependent addition of the 70-gene signature to a competing risk model with clinicopathological factors and radiotherapy, in the entire cohort of 1,053 primary operable breast cancer cases treated at the Netherlands Cancer Institute, and in subgroups according to surgery and radiotherapy. NRI indices are shown separately for those expected to experience LRR (with LRR) and those expected to remain free from LRR (without LRR) in the evaluated time period (5 years since diagnosis: left (a, c, e); 10 years: right (b, d, f)), as well as the sum of these two components (overall); NRI categorical (a, b) summarizes reclassification over all risk categories (low, intermediate, or high LRR risk1),

123


NRI clinical (c, d) shows net reclassification improvement in the clinicopathological model’s intermediate risk group, and NRI continuous (e, f) shows net reclassification improvement based on predicted probabilities without categorization. Grey horizontal reference lines denote absence of net change in classification performance; Error bars are 95 % confidence intervals based on 2,000-fold bootstrap resampling (exclusion of 0 % NRI denotes statistical significance at the two-sided 5 % level). 1Definition of low LRR risk: \2.5 % at 5 years and \5 % at 10 years; high risk: C7.5 % at 5 years and C15 % at 10 years; intermediate risk lies between these thresholds


609

Table 3 Discriminative ability of the locoregional recurrence prognostic models at 5 and 10 years following diagnosis in the entire cohort of 1,053 primary operable breast cancer cases treated at the Time since diagnosis model

Entire cohort

Netherlands Cancer Institute, and full models’ performance in subgroups according to surgery and radiotherapy.a

Breast-conserving therapy

Mastectomy without radiotherapy

Mastectomy with radiotherapy

C-index (95 % CI)

Pb

C-index (95 % CI)

Pb

C-index (95 % CI)

Pb

C-index (95 % CI)

P2

0.760 (0.700–0.820) 0.773 (0.717–0.830)

Ref.

0.805 (0.722–0.887) 0.822 (0.739–0.905)

Ref.

0.759 (0.657–0.860) 0.743 (0.645–0.841)

Ref.

0.733 (0.602–0.864) 0.772 (0.660–0.884)

Ref.

Through 5 years Traditional predictors 70-GS extended 70-GS extended, timedependent

0.13

0.24

0.37

0.019

0.776 (0.723–0.829)

0.30

0.822 (0.739–0.904)

0.47

0.729 (0.634–0.824)

0.31

0.796 (0.701–0.890)

0.036

Traditional predictors

0.731 (0.680–0.782)

Ref.

0.744 (0.670–0.817)

Ref.

0.757 (0.668–0.846)

Ref.

0.683 (0.560–0.805)

Ref.

70-GS extended

0.741 (0.693–0.790)

0.23

0.759 (0.688–0.831)

0.27

0.738 (0.652–0.825)

0.25

0.714 (0.603–0.826)

0.079

70-GS extended, timedependent

0.744 (0.6960.791)

0.54

0.763 (0.693–0.834)

0.53

0.724 (0.639–0.809)

0.16

0.733 (0.629–0.838)

0.11

Through 10 years

Number of patients in groupc

1,053

481

257

298

Number of LRR events observed3

87

39

25

20

70-GS 70-gene signature, CI confidence interval, LRR locoregional recurrence, Ref reference a

Subgroup C-indices were based on the predicted probabilities derived from the models fitted in the entire cohort but observed in the respective subgroups

b

P-values for change in C-index were obtained using 2,000-fold bootstrap resampling. 3Numbers do not add due to missing data

safely omitted. To allow such external validation, we have provided the formulas necessary to calculate a patient’s predicted LRR risk in the supplementary data (see Appendix 2, also including a formula adjusted for modeloverfitting induced by a relative paucity of LRR events in relation to the number of predictors). Two other studies used a gene expression classifier to predict LRR risk, both following BCT. Mamounas et al., with data from the NSABP B-14 and B-20 trials (895 patients, 73 LRR events) [26], observed a 10-year LRR risk of 15.8 % for patients with a high-risk 21-gene recurrence score after treatment with tamoxifen and 4.3 % among patients with a low-risk recurrence score. The Eastern Cooperative Oncology Group E2197 study (388 patients, 30 LRR events) [27] observed 10-year LRR rates in the 21-gene recurrence score high-risk group of 8.7 and 3.7 % in the low-risk group. Our univariable results are in line with these studies (with a 10-year LRR risk of 13.2 and 5.8 % for 70-gene signature high- and low-risk BCT-treated patients). Importantly, the NSABP study included 227 patients used to develop the 21-gene recurrence score, likely leading to overoptimistic results. The novelty in our

study is that our data allow additional analyses of patients treated with or without radiotherapy after mastectomy. Furthermore, in our study, the value of the 70-gene signature is evaluated in the context of standard clinicopathological prognostic factors as would be practiced in clinical medicine. In our study, addition of the 70-gene signature to standard clinicopathological factors is especially promising in the years directly following diagnosis and treatment for the group of patients treated with BCT. In those patients at intermediate 5-year LRR risk according to standard clinicopathological predictors, 22 % were reclassified downwards by the 70-gene signature with a substantial and significant net improvement in classification. Radiotherapy reduces the risk of LRR by at least 50 % in all patients, irrespective of patient and tumour characteristics or other treatment factors [3]. Considering the relative risk reduction of radiotherapy, LRR risk in the BCT group (whom all receive radiotherapy by default) is likely to have been at most 5 % at 5 years without radiotherapy. This is a generally accepted risk [23]. Applying a strategy to our data of withholding radiotherapy, if the

123

610 Table 4 Reclassification of 5-year locoregional recurrence (LRR) risk following the time-dependent addition of the 70-gene signature to a competing risk model with clinicopathological factors and

123

Breast Cancer Res Treat (2014) 148:599–613 radiotherapy, as observed in 1,053 primary operable breast cancer cases treated at the Netherlands Cancer Institute.a

Breast Cancer Res Treat (2014) 148:599–613 Table 5 Reclassification of 5-year locoregional recurrence (LRR) risk following the time-dependent addition of the 70-gene signature to a competing risk model with clinicopathological factors, as observed

611 in 481 primary operable breast cancer cases treated with breast conserving therapy at the Netherlands Cancer Institute.a

123

612

predicted 5-year risk of LRR is below 2.5 %, would have prevented radiotherapy in 22 % of those BCT-treated patients at intermediate risk according to clinicopathological factors, when the 70-gene signature would have been used in adjunct. This corroborates the recent suggestion that radiotherapy might be of limited value in some patients treated with BCT [28]. Our data further suggest limited value of radiotherapy in women aged 70 years or older at diagnosis. In conclusion, the 70-gene signature is a strong independent prognostic factor for LRR and shows high promise for added value in prognostication of LRR risk in breast cancer patients. The 70-gene signature may allow withholding radiation in low-risk patients treated with breastconserving surgery and also help differentiate which patients would benefit from radiation after mastectomy. Acknowledgments No financial support was received to perform this study. We acknowledge the enormous efforts of M. van de Vijver, M. Buyse, P. Bedard, J. Bueno-de-Mesquita, S. Mook, M. Kok, M. Saghatchian, and colleagues to perform the various 70-gene signature studies included in this study. We thank N. Russell for her input on the interpretation of the preliminary data. We especially thank the data-managers at the Netherlands Cancer Institute for all their efforts in collection of the follow-up data. Conflict of interest LvtV is named inventor on the patent for the 70-gene signature used in this study. LvtV reports being shareholder in and part-time employed by Agendia Inc, the commercial company that markets the 70-gene signature as MammaPrintTM. LvtV was supported by the Dutch Genomics Initiative ‘Cancer Genomics Centre’. FdS is director medical affairs of Agendia Inc. HB is a non-remunerated, nonstake holding member of the supervisory board of Agendia Inc. All other authors declare that they have no conflict of interest.

References 1. Canavan J, Truong PT, Smith SL et al (2014) Local recurrence in women with stage I breast cancer: declining rates over time in a large, population-based cohort. Int J Radiat Oncol Biol Phys 88:80–86. doi:10.1016/j.ijrobp.2013.10.001 2. van der Heiden-van der Loo M, Ho VKY, Damhuis RAM (2010) Percentage of local recurrence following treatment for breast cancer is not a suitable performance indicator. Ned Tijdschr Geneeskd 154:A1984 3. Early Breast Cancer Trialists’ Collaborative Group (EBCTCG), Darby S, McGale P, et al. (2011) Effect of radiotherapy after breast-conserving surgery on 10-year recurrence and 15-year breast cancer death: meta-analysis of individual patient data for 10,801 women in 17 randomised trials. Lancet 378:1707–1716. doi:10.1016/S0140-6736(11)61629-2 4. Nijenhuis MV, Rutgers EJT (2013) Who should not undergo breast conservation? Breast 22(Suppl 2):S110–S114. doi:10. 1016/j.breast.2013.07.021 5. Van’t Veer LJ, Paik S, Hayes DF (2005) Gene expression profiling of breast cancer: a new tumor marker. J Clin Oncol 23:1631–1635. doi:10.1200/JCO.2005.12.005 6. van t Veer LJ, Dai H, van de Vijver MJ (2002) Gene expression profiling predicts clinical outcome of breast cancer. Nature 415:530–536. doi:10.1038/415530a

123

Breast Cancer Res Treat (2014) 148:599–613 7. Drukker CA, Bueno-de-Mesquita JM, Rete`l VP et al (2013) A prospective evaluation of a breast cancer prognosis signature in the observational RASTER study. Int J Cancer 133:929–936. doi:10.1002/ijc.28082 8. van de Vijver MJ, He YD, Van’t Veer LJ et al (2002) A geneexpression signature as a predictor of survival in breast cancer. N Engl J Med 347:1999–2009. doi:10.1056/NEJMoa021967 9. Bueno-de-Mesquita JM, Linn SC, Keijzer R et al (2009) Validation of 70-gene prognosis signature in node-negative breast cancer. Breast Cancer Res Treat 117:483–495. doi:10.1007/ s10549-008-0191-2 10. Buyse M, Loi S, van’t Veer L et al (2006) Validation and clinical utility of a 70-gene prognostic signature for women with nodenegative breast cancer. J Natl Cancer Inst 98:1183–1192. doi:10. 1093/jnci/djj329 11. Wapnir IL, Anderson SJ, Mamounas EP et al (2006) Prognosis after ipsilateral breast tumor recurrence and locoregional recurrences in five national surgical adjuvant breast and bowel project node-positive adjuvant breast cancer trials. J Clin Oncol 24:2028–2037. doi:10.1200/JCO.2005.04.3273 12. Bueno-de-Mesquita JM, van Harten WH, Rete`l VP et al (2007) Use of 70-gene signature to predict prognosis of patients with node-negative breast cancer: a prospective community-based feasibility study (RASTER). Lancet Oncol 8:1079–1087. doi:10. 1016/S1470-2045(07)70346-7 13. Mook S, Schmidt MK, Viale G et al (2009) The 70-gene prognosis-signature predicts disease outcome in breast cancer patients with 1-3 positive lymph nodes in an independent validation study. Breast Cancer Res Treat 116:295–302. doi:10.1007/s10549-0080130-2 14. Saghatchian M, Mook S, Pruneri G et al (2013) Additional prognostic value of the 70-gene signature (MammaPrint(Ò)) among breast cancer patients with 4–9 positive lymph nodes. Breast. doi:10.1016/j.breast.2012.12.002 15. Kok M, Koornstra RH, Mook S et al (2012) Additional value of the 70-gene signature and levels of ER and PR for the prediction of outcome in tamoxifen-treated ER-positive breast cancer. Breast 21:769–778. doi:10.1016/j.breast.2012.04.010 16. Mook S, Schmidt MK, Weigelt B et al (2010) The 70-gene prognosis signature predicts early metastasis in breast cancer patients between 55 and 70 years of age. Ann Oncol 21:717–722. doi:10.1093/annonc/mdp388 17. Glas AM, Floore A, Delahaye LJMJ et al (2006) Converting a breast cancer microarray signature into a high-throughput diagnostic test. BMC Genomics 7:278. doi:10.1186/1471-2164-7-278 18. Wolbers M, Koller MT, Witteman JCM, Steyerberg EW (2009) Prognostic models with competing risks: methods and application to coronary risk prediction. Epidemiology 20:555–561. doi:10. 1097/EDE.0b013e3181a39056 19. Aalen O (1978) Nonparametric estimation of partial transition in multiple decrement models. Ann Stat 6:534–545 20. Gray RJ (1988) A class of K-samples tests for comparing the cumulative incidence of a competing risk. Ann Stat 16: 1141–1154 21. Fine JP, Gray RJ (1999) A proportional hazards model for the subdistribution of a competing risk. J Am Stat Assoc 94:496–509 22. Pencina MJ, D’Agostino RB, Steyerberg EW (2011) Extensions of net reclassification improvement calculations to measure usefulness of new biomarkers. Stat Med 30:11–21. doi:10.1002/sim. 4085 23. Rutgers EJ, EUSOMA Consensus Group (2001) Quality control in the locoregional treatment of breast cancer. Eur J Cancer 37:447–453 24. Clarke M, Collins R, Darby S et al (2005) Effects of radiotherapy and of differences in the extent of surgery for early breast cancer on local recurrence and 15-year survival: an overview of the

Breast Cancer Res Treat (2014) 148:599–613 randomised trials. Lancet 366:2087–2106. doi:10.1016/S01406736(05)67887-7 25. Donders ART, van der Heijden GJMG, Stijnen T, Moons KGM (2006) Review: a gentle introduction to imputation of missing values. J Clin Epidemiol 59:1087–1091. doi:10.1016/j.jclinepi. 2006.01.014 26. Mamounas EP, Tang G, Fisher B et al (2010) Association between the 21-gene recurrence score assay and risk of locoregional recurrence in node-negative, estrogen receptor-positive breast cancer: results from NSABP B-14 and NSABP B-20. J Clin Oncol 28:1677–1683. doi:10.1200/JCO.2009.23.7610

613 27. Solin LJ, Gray R, Goldstein LJ et al (2012) Prognostic value of biologic subtype and the 21-gene recurrence score relative to local recurrence after breast conservation treatment with radiation for early stage breast carcinoma: results from the eastern cooperative oncology group E2197 study. Breast Cancer Res Treat 134:683–692. doi:10.1007/s10549-012-2072-y 28. Tinterri C, Gatzemeier W, Costa A et al (2013) Breast-conservative surgery with and without radiotherapy in patients aged 55–75 years with early-stage breast cancer: a prospective, randomized, multicenter trial analysis after 108 months of median follow-up. Ann Surg Oncol. doi:10.1245/s10434-013-3233-x

123