Introduction Breast cancer is a heterogeneous disease that can be classified using a variety of clinical and pathological features. Classification may help in prognostication and targeting of treatment to those most likely to benefit. Currently, estrogen receptor (ER) status and human epidermal growth factor receptor-2 (HER2) status are routinely used as predictive markers to select specific adjuvant therapies. Prognostic markers may also be used to target adjuvant chemotherapy to those at highest risk of poor outcome—for example, the risk prediction tool Adjuvant!Online (www.adjuvant.org) uses prognostic markers to predict the likely absolute benefit of postoperative hormonal and/or chemotherapy and is widely used by oncologists to identify patients most likely to benefit from adjuvant treatment. Perou et al. identified four breast cancer subtypes on the basis of gene-expression profiling of 39 invasive breast tumours and three normal breast specimens [1]. There was one ER-positive (ER+/luminal-like) and three ER-negative subtypes (basal-like, ERBB2+, and normal-like). In addition to expressing the ER receptor, luminal-like tumours expressed other genes that were characteristic of luminal or glandular epithelial cells of origin. The basal-like tumours expressed basal or myoepithelial markers, and none of the basal tumours expressed ER. Similar to the basal-like tumours, overexpression of the ERBB2 oncogene was associated with low ER. The normal-like subgroup was typified by high gene expression for basal and low expression for luminal breast epithelium. A subsequent gene expression analysis by Sorlie et al. of patterns in 78 breast cancers, three fibroadenomas, and four normal breast tissues suggested that the luminal-like subtype could be further separated into two subgroups: luminal A and luminal B [2]. The molecular subtypes were reflected in differences in prognosis. Overall and relapse-free survivals were most favourable for luminal A tumours and least favourable for ERBB2+ and basal-like breast cancers. The investigators also suggested that there may be a third luminal subgroup, the luminal C tumours, but this has not been supported by the subsequent analysis of an expanded dataset [3]. The classification of breast cancers into subgroups on the basis of gene expression patterns in tumour tissue is often regarded as the gold standard, but widespread use of gene-expression profiling in either the clinical or the research setting remains limited. Lack of widespread use of expression profiles is primarily due to the expense and technical difficulty encountered when carrying out high-throughput gene-expression profiling using paraffin-embedded material. Moreover, the currently defined subtypes based on expression profiling were determined through the study of relatively small numbers of tumours and these subgroups may not be definitive. Consequently there is interest in using immunohistochemical (IHC) markers to classify tumours into subtypes that are surrogates for those based on gene-expression profiling [4]. Many investigators have used IHC to classify tumours but have used different naming conventions. Generally a hierarchical classification is used, with luminal and nonluminal tumours defined as those tumours that express either ER or progesterone receptor (PR) and those that do not. The luminal and nonluminal groups can then be further subdivided according to HER2-expression status to generate four subtypes, and these four subtypes can each be categorised according to whether or not they express a basal marker yielding a total of eight subtypes. The mapping of these eight IHC subtypes onto the five subtypes based on gene expression is not exact. Luminal A tumours as defined by gene expression have, in general, higher expression of ER-related genes and lower expression of proliferative genes than luminal B tumours [5]. However, there are no established IHC markers for subdividing the luminal subtypes into the same categories. Recently, it has been suggested that the luminal B subtype is equivalent to those that express either HER2 or the proliferation marker KI67 [6]. The nonluminal tumours are ER negative and PR negative and are generally subdivided into three groups. The nonluminal, HER2-positive tumours are the equivalent of the ERBB2-overexpressing tumours. Tumours that do not express ER, PR, or HER2—the triple negative phenotype (TNP) tumours—are often regarded as equivalent to the basal subtype as they can be easily identified with IHC markers that are currently used in routine clinical use. However, not all TNP tumours express basal cytokeratins (CKs), and within the TNP subtype, expression of basal markers may reflect important clinical differences. Expression of either CK5/6 or epidermal growth factor receptor (EGFR) has been shown to accurately identify basal-like tumours classified using gene expression [7],[8], and several published studies have used these markers to subclassify the TNP tumours into a core basal subgroup (CBP), which is equivalent to the basal-like from expression profiling and the five negative phenotype (5NP: ER−, PR−, HER2−, CK5/6−, and EGFR−). Although this hierarchical classification is commonly used, questions remain as to whether these groups are biologically distinct and clinically relevant. For example, it has been suggested that basal markers can be used to classify the basal tumours independent of other markers [9]. Cheang et al. reported a significantly poorer survival in CBP tumours compared to the 5NP tumours [10], an observation that supports the notion that the two are biologically distinct types of the TNP tumours. This finding was not confirmed by a smaller study with limited power to detect small differences [11]. A third study reported that the prognostic significance of CBP tumours was similar to that of the TNP tumours [12]. However, they did not explicitly compare the CBP and 5NP subtypes. Previously published studies have either compared the five subtypes by using the luminal HER2-negative tumours as a reference category to compare with the other four subtypes [8],[10],[12],[13], or they have compared the subtypes by restricting the analysis to either luminal or nonluminal tumours [6],[11]. Unanswered questions include whether the behaviour of luminal HER2-positive tumours and the nonluminal HER2-positive tumours are different, whether the behaviour of luminal basal-positive tumours is different from that of the nonluminal basal-positive tumours, and whether basal marker status is important in the luminal, HER2-negative tumours. The association between ER status and mortality is known to be time dependent, with hazard ratios for ER-positive versus ER-negative tumours being lower than one in the first years after diagnosis and becoming higher than one after 7–10 y. Mortality in women with ER-positive tumours remains fairly constant over time, whereas the mortality in women with ER-negative tumours is initially higher than that in women with ER-positive disease and then falls to a lower rate after 7–10 y [14]–[16]. In addition, Tischkowitz and colleagues reported that the prognostic effects of both TNP and CBP tumours compared to luminal tumours tended to diminish over time, whereas the effect of CK5 and other basal markers, when considered alone, might increase with time [12]. Another study reported that the effects of the CBP were attenuated over time [13]. Inspection of the Kaplan-Meier survival curves published by Cheang et al. also suggest that the prognostic effects of the CBP and 5NP subtypes are time dependent [10]. All the major subtypes apart from the luminal A tumours are relatively infrequent, and only very large studies with prolonged follow-up have the power to study meaningful differences in prognosis. The aim of this study was to pool individual data from multiple breast cancer case series, in order to definitively establish the relative survival of the major subtypes of breast cancer as classified using five IHC markers, and to characterise their prognostic effects over time. Materials and Methods Ethics Statement All studies were approved by the relevant research ethics committee or institutional review board. Participants in Amsterdam Breast Cancer Study (ABCS), Helsinki Breast Cancer Study (HEBCS), Jewish General Hospital (JGH), Mayo Clinic Breast Cancer Study (MCBCS), Melbourne Collaborative Cohort Study (MCCS), Polish Breast Cancer Study (PBCS), Sheffield Breast Cancer Study (SBCS), and Study of Epidemiology and Risk factors in Cancer Heredity (SEARCH) provided informed written consent. Samples for British Columbia Cancer Agency (BCCA), Nottingham Breast Cancer Case Series (NOBCS), University of British Columbia (UBC), and Vancouver General Hospital (VGH) were from legacy archival material and individual consent was not obtained. All data were anonymised before being sent to the coordinating centre for analysis. Study Populations The international breast cancer association consortium (BCAC) comprises a large number of studies investigating the role of common germline genetic variation in breast cancer susceptibility [17]. In addition to data on germline genotype, many BCAC studies have detailed pathological data on the breast cancer cases linked to follow-up data. All BCAC studies that had collected IHC data on ER, PR, HER2, and either EGFR or CK5/6 or both, in addition to survival time data and data on tumour grade, size, and nodal status were eligible for inclusion in this study. The investigators of the three previously published studies with equivalent data [10]–[12], were also invited to contribute their data, as were the investigators of a fourth large breast cancer case series that had taken part in a previous collaboration involving other BCAC studies [18]. All studies provided data on age at diagnosis, vital status, breast cancer-specific mortality, time between diagnosis and ascertainment, follow-up time, tumour grade (low, intermediate, and high), tumour size (<2 cm, 2–4.9 cm, ≥5 cm) and node status (positive or negative). In total, 12 studies from Europe, North America, and Australia contributed data on 10,159 cases with complete data [7],[9],[10],[12],[18]–[29]. Nine studies also provided data on whether or not the patient had been treated with adjuvant hormonal therapy or adjuvant chemotherapy. These data were available for a subset of 8,171 and 8,061 cases, respectively. The studies are described in Table 1. 10.1371/journal.pmed.1000279.t001 Table 1 Description of participating studies. Study Country Case Ascertainment Case Definition Age Range (y) References ABCS The Netherlands Hospital-based All cases of operable, invasive cancer diagnosed from 1974 to 1994 in four Dutch hospitals. Familial non-BRCA1/2 cases <50 from the Clinical Genetic Centre at The Netherlands Cancer Institute 23–50 [19] BCCA Canada Hospital-based Women diagnosed with invasive breast cancer between 1986 to 1992 and identified through the British Columbia Cancer Agency 23–89 [7],[10] HEBCS Finland Hospital-based (1) Consecutive cases (883) from the Department of Oncology, Helsinki University Central Hospital 1997–1998 and 2000; (2) Consecutive cases (986) from the Department of Surgery, Helsinki University Central Hospital 2001–2004; (3) Familial breast cancer patients (536) from the Helsinki University Central Hospital, Departments of Oncology and Clinical Genetics (1995–) 22–96 [20]–[22] JGH Canada Hospital-based Ashkenazi Jewish women diagnosed with nonmetastatic, invasive breast cancer at Jewish General Hospital, Montreal between 1980 and 1995 26–66 [12] MCBCS USA Hospital-based Incident cases residing in six states (Minnesota, Wisconsin, Iowa, Illinois, North Dakota, South Dakota) seen at the Mayo Clinic in Rochester, Minnesota from 2002–2005 22–89 [23] MCCS Australia Cohort Incident cases diagnosed within the Melbourne Collaborative Cohort Study during the follow-up from baseline (1990–1994) to 2004 of the 24,469 participating women 30–82 [24] NOBCS UK Hospital-based Primary operable breast carcinoma patients presenting from 1986 to 1998 and entered into the Nottingham Tenovus Primary Breast Carcinoma Series. 26–93 [9] PBCS Poland Population-based Incident cases from 2000–2003 identified through a rapid identification system in participating hospitals covering ∼90% of all eligible cases; periodic check against the cancer registries in Warsaw and Łódź to assure complete identification of cases 27–75 [25] SBCS UK Hospital-based Women with pathologically confirmed breast cancer recruited from surgical outpatient clinics at the Royal Hallamshire Hospital, Sheffield, 1998–2002; cases are a mixture of prevalent and incident disease 29–93 [26],[27] SEARCH UK Population-based Two groups of cases identified through East Anglian Cancer Registry: (1) prevalent cases diagnosed age <55 y from 1991–1996 and alive when study started in 1996; (2) incident cases diagnosed age <70 y diagnosed after 1996 23–69 [18] UBCBCT Canada Hospital-based Women with stage I to III breast cancer who participated in four different British Columbia Cancer Agency clinical trials between 1970 and 1990 and all received chemotherapy 22–90 [28],[29] VGH Canada Hospital-based Women with primary breast cancer who underwent surgery at Vancouver General Hospital 1975–1995 28–91 [12] ABCS, Amsterdam Breast Cancer Study; BCCA, British Columbia Cancer Agency; HEBCS, Helsinki Breast Cancer Study; JGH, Jewish General Hospital; MCBCS, Mayo Clinic Breast Cancer Study; MCCS, Melbourne Collaborative Cohort Study; NOBCS, Nottingham Breast Cancer Case Series; PBCS, Polish Breast Cancer Study; SBCS, Sheffield Breast Cancer Study; SEARCH, Study of Epidemiology and Risk factors in Cancer Heredity; UBCBCT, University of British Columbia Breast Cancer Trials; VGH, Vancouver General Hospital. Immunohistochemistry and Tumour Classification Data for these antibodies were either derived from IHC performed in a research setting or collated from patient records by the individual groups. The methods used by each study for each marker are shown in Table S1. The cases were grouped into subtypes on the basis of their protein expression profile (Figure 1). Luminal tumours were those with positive staining for ER or PR. Luminal tumours were subdivided according to HER2 status into luminal 1 (HER2-negative), which is broadly equivalent to the luminal A tumours defined by gene expression, and luminal 2 (HER2-positive) tumours. The luminal 2 tumours are a subset of the luminal B tumours because some of the tumours classified as luminal 1 would be expected to express proliferative markers and thus be misclassified luminal B tumours. The nonluminal tumours were those that were negative for both ER and PR. These were subdivided by HER2 expression status into the nonluminal HER2-positive tumours and the TNP tumours. The TNP tumours were further subdivided into the CBP tumours (either CK5/6 or EGFR positive) and the 5NP tumours (CK5/6-negative and EGFR-negative). Four studies did not provide data for EGFR, and for these studies the 5NP tumours were those that were negative for ER, PR, HER2, and CK5/6. A small number of 5NP tumours from these studies will thus be misclassified core basal tumours. The tumours classified as luminal 1 were also further subdivided according to expression of basal markers into luminal 1, basal marker negative and luminal 1, basal marker positive. 10.1371/journal.pmed.1000279.g001 Figure 1 Classification of breast cancer subtypes according to IHC marker profile. Statistical Analysis The association between each prognostic marker and subtype and all-cause mortality after diagnosis was investigated using Cox regression stratified by study and adjusted for age at diagnosis, grade, node status, and size of tumour. Ordinal categories of tumour grade and size were treated as continuous variables in all analyses. Age at diagnosis was treated as a categorical variable (<40, 40–49, 50–59, and ≥60 y). In several studies the cases were ascertained after diagnosis (prevalent cases), and this was allowed for in the analysis by setting “time at risk” from the date of diagnosis and “time under observation” on date of study entry. This step produces an unbiased estimate of the hazard ratio provided the proportional hazards assumption is correct [16]. Follow-up was censored on the date of death from any cause, or, if death did not occur, on date last known alive or at 15 y after diagnosis, whichever came first. The Cox proportional hazards model assumes that the hazard ratio is constant over time. This assumption is known to be violated for ER [14]–[16] and over prolonged follow-up is also likely to be violated for other predictors. We therefore carried out a conditional relative survival analysis by splitting follow-up time into five different periods—0–2, 2–4, 4–6, 6–10, and 10–15 y after diagnosis—and deriving Cox models separately for each period. The Cox proportional hazards assumption was checked for each study period by visual inspection of the standard log-log plots. A test for heterogeneity of the study-specific hazard ratios was carried out using the Mantel-Haenszel method. Kaplan-Meier cumulative survival plots were adjusted for study, age group, tumour grade, tumour size, and node status. In order to provide an overall test of association to compare survival time across all 15 y of follow-up we used multivariate Cox regression models in which the prognostic factors were treated as time-varying covariates. In these models the log hazard ratio varies as a function of the natural logarithm of follow-up time. Models with and without the covariates of interest were then compared using likelihood ratio tests. All analyses were performed in Intercooled Stata, version 10 (Stata Corp). Results Eight studies provided data on ER, PR, HER2, CK5/6, and EGFR with a further four studies providing data on ER, PR, HER2, and CK5/6, but not EGFR. Based on these data, there were 10,159 subjects that could be classified into one of the five major breast subtypes. There were 3,181 deaths in 85,799 person-years of follow-up, with 1,975 deaths from breast cancer. The multivariate, period-specific hazard ratios for age (in four categories), tumour grade, tumour size, node status, and the IHC markers are given in Table 2. These data show that the hazard ratios for all variables except age at diagnosis attenuate over time, and that for ER, PR, HER2, CK5/6, EGFR, and grade the effect changes direction with time. The time-dependent changes were most pronounced for ER and PR status. There was little difference in the hazard ratios for all-cause mortality and breast cancer-specific mortality, except for in the youngest and oldest age groups (Figures S1 and S2). Breast cancer-specific hazard ratios tended to be higher for women diagnosed under the age of 40 y (reference age at diagnosis 50–59 y). In contrast, for age at diagnosis ≥60 y, all-cause mortality hazard ratios were greater, as might be expected because of the impact of mortality from other causes. 10.1371/journal.pmed.1000279.t002 Table 2 Multivariate period-specific all-cause mortality hazard ratios (95% CI). Variable Time after Diagnosis 0–2 y 2–4 y 4–6 y 6–10 y 10–15 y Age at diagnosis (y) <40 0.69 (0.49–0.98) 1.09 (0.87–1.37) 1.14 (0.84–1.55) 0.83 (0.62–1.12) 0.68 (0.44–1.05) 40–49 0.63 (0.48–0.84) 0.77 (0.64–0.93) 0.83 (0.65–1.06) 0.66 (0.53–0.82) 0.51 (0.38–0.68) 50–59 1.00 (ref) 1.00 (ref) 1.00 (ref) 1.00 (ref) 1.00 (ref) ≥60 1.74 (1.36–2.22) 1.26 (1.04–1.52) 1.64 (1.31–2.06) 1.79 (1.49–2.14) 2.05 (1.63–2.58) Grade a 1.51 (1.24–1.84) 1.81 (1.59–2.08) 1.37 (1.18–1.60) 1.14 (1.01–1.29) 0.97 (0.83–1.13) Node positive 2.64 (2.12–3.27) 2.42 (2.09–2.82) 1.86 (1.55–2.23) 1.56 (1.35–1.82) 1.40 (1.15–1.70) Tumour size a 1.67 (1.42–1.97) 1.47 (1.31–1.66) 1.43 (1.23–1.66) 1.37 (1.20–1.56) 1.30 (1.09–1.55) ER positive 0.55 (0.42–0.71) 0.76 (0.63–0.91) 1.31 (1.02–1.68) 1.63 (1.29–2.07) 1.24 (0.91–1.69) PR positive 0.36 (0.27–0.47) 0.62 (0.52–0.74) 0.74 (0.6–0.91) 1.04 (0.87–1.23) 1.16 (0.92–1.46) HER2 positive 1.21 (0.95–1.52) 1.50 (1.27–1.78) 1.55 (1.23–1.96) 1.35 (1.07–1.69) 0.96 (0.67–1.37) Basal marker positive 1.33 (1.06–1.68) 1.21 (1.01–1.44) 1.38 (1.08–1.78) 1.06 (0.83–1.35) 0.83 (0.59–1.17) All analyses are stratified by study. a Grade and tumour size are ordinal variables treated as continuous, giving hazard ratios per unit increase in score. There were 7,882 luminal tumours (78% of total). Of these, 7,243 (92%) were luminal 1 and 639 (8%) were luminal 2. There were 632 tumours of the nonluminal HER2-positive subtype (6% of total), and 1,645 TNP tumours (16% of total). Of the TNP tumours, 962 were CBP (58%) and 683 basal-negative tumours (42%). The number of tumours by the five major subtypes for each study are shown in Table 3. In addition to the five main subtypes, we subdivided the luminal 1 tumours according to expression of basal markers, with 562 (8%) being basal marker positive and 6,119 (92%) being basal marker negative (Table S2 shows the luminal 1 subgroups by study). Table 4 shows the characteristics of the five major breast cancer subtypes by age at diagnosis, tumour grade, tumour size, and node status. 10.1371/journal.pmed.1000279.t003 Table 3 Number of tumours by subtype and study. Study Luminal 1 Luminal 2 Nonluminal HER2+ CBP 5NP Total n Percent n Percent n Percent n Percent n Percent ABCS 497 67 64 9 51 7 60 8 68 9 740 BCCA 2,378 71 206 6 238 7 317 9 209 6 3,348 HEBCS 169 72 25 11 8 3 21 9 13 6 236 JGH 160 77 18 9 5 2 21 10 3 1 207 MCBCS 219 86 24 9 4 2 8 3 1 <1 256 MCCS 276 72 22 6 30 8 37 10 17 4 382 NOBCS 1,051 71 44 3 71 5 196 13 108 7 1,470 PBCS 694 69 35 3 67 7 137 14 75 7 1,008 SBCS 206 77 16 6 10 4 14 5 21 8 267 SEARCH 1,247 76 121 7 71 4 112 7 83 5 1,634 UBC 154 42 53 15 62 17 15 4 81 22 365 VGH 192 78 11 4 15 6 24 10 4 2 246 Total 7,243 71 639 6 632 6 962 9 683 7 10,159 ABCS, Amsterdam Breast Cancer Study; BCCA, British Columbia Cancer Agency; HEBCS, Helsinki Breast Cancer Study; JGH, Jewish General Hospital; MCBCS, Mayo Clinic Breast Cancer Study; MCCS, Melbourne Collaborative Cohort Study; NOBCS, Nottingham Breast Cancer Case Series; PBCS, Polish Breast Cancer Study; SBCS, Sheffield Breast Cancer Study; SEARCH, Study of Epidemiology and Risk factors in Cancer Heredity; UBCBCT, University of British Columbia Breast Cancer Trials; VGH, Vancouver General Hospital. 10.1371/journal.pmed.1000279.t004 Table 4 Characteristics of breast cancer subtypes by age at diagnosis, tumour grade, tumour size, and node status. Breast Cancer Subtype Characteristics Luminal 1 Luminal 2 HER2-enriched CBP 5NP Total n Percent n Percent n Percent n Percent n Percent n Percent Vital status at censoring Alive 5,242 72 369 58 333 53 590 61 444 65 6,978 69 Dead 2,001 28 270 42 299 47 372 39 239 35 3,181 31 Age group (y) <40 457 6 74 12 80 13 165 17 90 13 866 9 40–49 1,960 27 215 34 190 30 286 30 237 35 2,888 28 50–59 3,142 43 233 36 268 42 377 39 238 35 4,258 42 ≥60 1,684 23 117 18 94 15 134 14 118 17 2,147 21 Tumour grade 1 1,493 21 41 6 20 2 15 3 40 6 1,609 16 2 3,645 50 239 37 146 23 129 13 174 25 4,333 42 3 2,105 29 359 56 466 73 818 85 469 69 4,217 42 Node status Negative 4,229 58 278 44 267 42 577 60 367 54 5,718 56 Positive 3,014 42 361 56 365 58 385 40 316 46 4,441 44 Tumour size <2 cm 4,441 61 300 47 272 43 442 46 296 43 5,751 56 2–4.9 cm 2,580 36 306 48 318 50 468 49 336 49 4,008 39 ≥5 cm 222 3 33 5 42 7 52 5 51 7 402 4 The hazard ratios over time for the five subtypes of breast cancer, stratified by study and adjusted for grade, tumour size, and node status, are shown in Figure 2. There was little evidence for heterogeneity of effects by study for these hazard ratios except for the 5NP tumours (Table S3). Figure 2 shows that, compared to the luminal 1 tumours, luminal 2 tumours are associated with a slightly poorer prognosis in the first few years after diagnosis, but that the difference reduces with time, and by 8 y after diagnosis there is no difference between the two. In contrast the mortality for women with the HER2-enriched and both types of TNP tumours (CBP and 5NP) is substantially greater than that for women with the luminal 1 tumours immediately after diagnosis, but the difference declines rapidly and reverses at 5–10 y after diagnosis. These patterns reflect the time-dependent changes in mortality rates in the different subgroups (Figure S3). Within the TNP subgroup, the women with CBP tumours have a slightly poorer prognosis than women with the 5NP tumours. This difference declines slightly over time and by 8 y after diagnosis, no difference is observed. A similar pattern is seen for the luminal 1, basal-positive tumours when compared to the luminal 1, basal-negative tumours. We repeated the analyses using breast cancer-specific mortality as the end point (Figure S4). The hazard ratio estimates tended to be greater (for hazard ratios greater than unity) than the all-cause mortality hazard ratios, but the confidence intervals were somewhat wider. 10.1371/journal.pmed.1000279.g002 Figure 2 Period-specific hazard ratios (all-cause mortality) for major breast cancer subtypes. All hazard ratios are stratified by study and adjusted for tumour grade, tumour size, and node status. The Kaplan-Meier cumulative survival for the three luminal subtypes adjusted for study, grade, tumour size, and node status is shown in Figure 3A. This result shows that the cumulative survival for the luminal 1 subtypes declines almost linearly over time, which is compatible with a constant mortality rate. In contrast, the mortality rate in women with the luminal 2 tumours tends to flatten out over time as the high mortality in the first few years after diagnosis declines. It also clearly shows the poorer prognosis for the luminal 1 tumours that are basal marker positive. The survival curves associated with nonluminal HER2-positive, CBP, and 5NP tumours all show a similar pattern to that of the luminal 2 tumours (Figure 3B). There were significant differences in prognosis between all pairs of subtypes apart from the nonluminal HER2-positive tumours compared with the CBP tumours (Table S4). Of particular note is the difference between the CBP and 5NP tumours (p = 0.0008). The luminal, HER2-positive tumours and the nonluminal, HER2-positive tumours are two distinct subgroups, with the nonluminal tumours having a poorer prognosis (p<0.0001), and the CBP tumours having a poorer prognosis than the luminal, basal-positive tumours (p<0.0001). These differences did not depend on whether or not the patient had been treated with either adjuvant hormone therapy or adjuvant chemotherapy (Figure S5). In contrast, the basal markers seem to have no prognostic significance within the HER2 positive subtypes of disease (p = 0.85). 10.1371/journal.pmed.1000279.g003 Figure 3 (A and B) Kaplan-Meier cumulative survival (all-cause mortality) in luminal and nonluminal tumours by subtype. All curves are adjusted for age at diagnosis, tumour grade, tumour size, node status, and study. The luminal, HER2-positive tumours and the nonluminal, HER2-positive tumours represent two distinct subgroups, as do the ER-positive/negative tumours that are basal positive. In both cases the ER-negative tumours have a poorer prognosis in the first few years after diagnosis, but after 5 to 10 y it is the ER-positive tumours that have the poorer outcome (Figure S6). In contrast, the basal markers seem to have no prognostic significance within the HER2-positive subtypes of disease (unpublished data). Data on the association between the major subtypes and prognosis have previously been published for three of the studies included in this analysis—BCCA, JGH, and VGH—and it is possible that the effect estimates that we report here are subject to publication bias. We therefore repeated all the analyses after excluding the data for these three studies but there was little difference in the results (see Figure S7). Discussion We evaluated the prognostic significance of five previously described major subtypes of breast cancer that were classified using five IHC markers. To our knowledge, this study represents one of the largest datasets analysed for prognosis research in breast cancer using IHC markers. Our data confirm the observations of others that the pattern of survival in ER-positive tumours is qualitatively different to that in ER-negative tumours. In ER-positive tumours, the mortality rate is approximately constant over time since diagnosis, whereas the mortality rate associated with ER-negative disease is initially high and then progressively declines over time. However, the pattern of mortality rates associated with the HER2-positive subgroup of ER-positive tumours (luminal 2) is similar to those of the nonluminal subtypes (Figure 3A). Berry et al. suggest [14] that the pattern of mortality after diagnosis associated with ER-positive tumours is mainly an effect of treatment with adjuvant hormone therapy and that the pattern of mortality in women not treated with adjuvant hormone therapy is similar to that in women with ER-negative disease. The pattern of mortality in women with luminal 1 tumours and treated with adjuvant hormone therapy was similar to those who did not receive hormone therapy (Figure S3). This result implies that the time-dependent effects we observed are not simply the result of adjuvant hormone therapy in a subset of the women with ER-positive tumours. Few of the participants with HER2-positive tumours in this study would have been treated with trastuzumab and so the prognosis in women with these tumours would not reflect the benefit of targeted therapy. Instead we propose that the survival patterns reflect the underlying molecular heterogeneity of breast cancer. We have hypothesized that this heterogeneous biology reflects the fact that breast cancers can initiate in different cell types, either breast epithelial stem cells or their progeny (transit amplifying cells or committed differentiated cells) [30]. Furthermore the recognition of the subtype-specific differences in short-term and long-term prognosis will inevitably lead to tailored follow-up programmes after completion of primary therapy. Our data confirm the view that the TNP is not a good proxy for the CBP because the CBP and 5NP tumours are biologically distinct and show different behaviours. The CBP tumours are clearly associated with a poorer prognosis than the 5NP tumours. Currently, chemotherapy remains the only systemic treatment option available for patients with triple negative (CBP and 5NP) tumours. A number of small studies have shown that basal-like cancers defined through gene-expression profiling or immunophenotyping are responsive to chemotherapy regimes [31]–[33]. In addition, the expression of core basal markers such as EGFR, may lead to the application of targeted therapies, with EGFR inhibitors currently under investigation for use in basal-like breast cancers. We have also shown that the expression of basal markers in ER-positive tumours is associated with a poorer prognosis, suggesting that the luminal 1 tumours represent two distinct subtypes, both of which differ in behaviour from the luminal 2 tumours. Overall the prognostic model based on the six subtypes defined by five IHC markers fits significantly better than a model based on three subtypes—ER-positive or PR-positive and HER2-negative, HER2-positive, and triple-negative tumours—defined by the three markers currently in standard clinical practice (likelihood ratio chisq = 54.4, 3 degrees of freedom [df], p<0.0001). One remaining question is whether the 5NP tumours represent a distinct subtype or are just other subtypes that have been misclassified because of assay failure. However, given the pattern of mortality rates over time since diagnosis (Figure S3), it seems unlikely that many of the 5NP tumours are misclassified luminal tumours. If the 5NP tumours were misclassified nonluminal HER2-positive or CBP tumours, we would expect the survival associated with them to be intermediate, whereas the 5NP tumours have a better prognosis than both the other nonluminal subtypes. Furthermore, the prognosis associated with the 5NP is different from each of the other five subtypes and is also different from all the other subtypes combined. Thus it seems likely that the majority of 5NP tumours represent a true distinct subtype, with a small, but unknown, proportion representing misclassification of the other subtypes, Until a marker to positively identify the genuine 5NP subtype has been identified, it will not be possible to separate these two sets of tumours. Our study has several limitations. IHC was carried out in different laboratories using different methods for both staining and scoring and, as a result, some misclassification of tumour subtypes is inevitable. However, it is likely that such error is random with respect to patient outcome. For the analyses of breast cancer-specific mortality, cause of death was obtained from the underlying cause of death as reported on death certificates and may thus be associated with some error. However, any error in ascertaining cause of death is likely to be random with respect to tumour characteristics. Thus, measurement error of either breast cancer subtype, as a result of interlaboratory variability or outcome, is, if anything, likely to result in an underestimate of any true differences between subtypes. The fact that we have found clear differences in subtypes classified by IHC analyses that were carried out in different laboratories, and would therefore be subject to interlaboratory assay result variability, suggests that the markers are robust to interlaboratory variation in their application and therefore suitable for use in routine clinical practice. There is also some nonrandom error as the luminal 1 tumours that express proliferation markers are likely to behave more like luminal 2 tumours [6]. As the luminal 1 tumours were used as the reference category, this misclassification is likely to lead to an underestimation in the true difference between luminal 1 and the other subtypes. Similarly, some of the 768 5NP tumours will be misclassified CBP tumours because data on EGFR were missing. Assuming these data were missing at random, approximately 25 of the 5NP tumours may represent misclassified CBP tumours. However, when the definition of 5NP tumours was restricted to those that were negative for both CK5/6 and EGFR, there was little difference in the hazard ratio estimates (unpublished data. Finally, the effects may also be underestimated because of the nonrandom use of adjuvant chemotherapy. The more aggressive subtypes are more likely to have been treated with chemotherapy, which would result in a reduction in the difference between these groups and the better prognosis subtypes. Data from 12 different studies were used in this analysis. These studies represent different ethnic groups from different regions of the world as well as differences in case ascertainment. Furthermore there were differences in the way that pathology samples were handled, stained, and scored, and the degree of misclassification will vary from study to study. This heterogeneity in study design may weaken the observed associations, and limit the specificity of the conclusions drawn. Nevertheless, the clear differences between the subtypes of breast cancer that we identified, despite the presence of heterogeneity, make the results robust and broaden their generalisability. In conclusion, we have confirmed that six breast cancer subtypes can be robustly classified using five IHC markers. These subtypes behave differently with specific patterns of mortality over time since diagnosis. These characteristics are independent of other clinico-pathological markers of prognosis and independent of systemic therapy received. The classification based on these markers is robust to multiple sources of heterogeneity between studies suggesting that they are suitable for use in routine clinical practice. The incorporation of these markers into prognostic tools such as Adjuvant!Online and the Nottingham Prognostic Index currently used in clinical practice or tools such as PREDICT [34], which was recently developed to enable the incorporation of novel prognostic biomarkers, may be warranted. It is plausible that these markers are predictive and that different subtypes respond differently to specific treatments, and the evaluation of subtype-specific responses in the context of clinical trials of specific treatments is urgently required. Given that these subtypes can easily be defined using robust IHC markers in archival material, this type of analysis should be possible with existing clinical trial data. Supporting Information Figure S1 Comparison of multivariate, period-specific hazard ratios for age group, tumour grade, and node status based on all-cause and breast-specific mortality. Left-hand panel are results for all-cause mortality and right-hand panels results for breast-specific mortality. Tumour size was treated as an ordinal variable in the Cox regression models and so the hazard ratios represent the hazard ratio for a unit change in the variable. (1.29 MB EPS) Click here for additional data file. Figure S2 Comparison of multivariate, period-specific hazard ratios for tumour size, ER, PR, HER2, and basal marker status based on all-cause and breast-specific mortality. Left-hand panel are results for all cause mortality and right-hand panels results for breast specific mortality. Tumour size was treated as ordinal variables in the Cox regression models and so the hazard ratios represent the hazard ratio for a unit change in the variable. (1.26 MB EPS) Click here for additional data file. Figure S3 Breast cancer-specific mortality by subtype and time since diagnosis. (0.65 MB EPS) Click here for additional data file. Figure S4 Period-specific hazard ratios (breast-specific mortality) for major breast cancer subtypes. All hazard ratios are stratified by study and adjusted for tumour grade, tumour size, and node status. (1.01 MB EPS) Click here for additional data file. Figure S5 Kaplan-Meier cumulative survival in luminal and nonluminal tumours by subtype and by treatment with adjuvant hormone therapy and adjuvant chemotherapy. All curves are adjusted for age at diagnosis, tumour grade, tumour size, node status, and study. (2.18 MB EPS) Click here for additional data file. Figure S6 Period-specific hazard ratios for ER-negative versus ER-positive disease stratified by HER2 status and basal marker status. All hazard ratios are adjusted for age at diagnosis, tumour grade, tumour size, and node status and stratified by study. (0.81 MB EPS) Click here for additional data file. Figure S7 Comparison of period- and subtype-specific hazard ratios (all-cause mortality) for all data and for subset of data after excluding published studies. Left-hand panels show results based on all data (as shown in Figure 1) and right-hand panels show equivalent hazard ratios after exclusion of data from BCCA, JGH, and VGH. (1.24 MB EPS) Click here for additional data file. Table S1 Methods used for IHC analysis by study. (0.10 MB DOC) Click here for additional data file. Table S2 Classification of luminal 1 tumours by basal marker expression. (0.04 MB DOC) Click here for additional data file. Table S3 p-Values for test for heterogeneity of period-specific hazard ratio estimates (compared to luminal 1 tumours) by study. (0.03 MB DOC) Click here for additional data file. Table S4 Likelihood ratio test statistic (2 degrees of freedom) and p-value for comparison of 15-y all-cause mortality between each subtype pair. (0.04 MB DOC) Click here for additional data file.