Breast cancer is a heterogenous disease whose behaviour is determined by the molecular characteristics of the tumour. In clinical practice, we rely on clinico-pathological features to predict tumour behaviour and patient outcome. These are powerful independent prognosticators 12 but are imperfect and represent only crude measures of the biological behaviour of a tumour. The power of these factors can be increased when they are used in combination e.g. the Nottingham Prognostic Index (NPI) 1 but they cannot predict outcome for all patients 345 and, as a result, many patients receive unnecessary treatment.

In the last 30 years, intensive efforts have been made to identify tools to improve prognostication. These range from biological markers to computer-based algorithms such as ADJUVANT! 6 that incorporate clinico-pathological and biological features. Early immunohistochemical studies identified a plethora of biological markers that can be triaged into categories based on the strength of evidence supporting their prognostic role 7. However, twenty years on only steroid hormone receptors and HER2 are used in clinical practice. The translation of these results has in part been impeded by a lack of robustness of the original research findings. Many studies were small and false positives are the result of a combination of low statistical power and publication bias. Furthermore, many prognostic markers are correlated and small studies have limited power to show independence of effects of novel markers in multivariate analyses.

Prognostic gene-based signatures have been reported with the claim that they out-perform current standards 89 although this has been questioned 101112. While promising, inconsistencies relating to quality control, tissue handling, sample size and data analysis need to be addressed before the clinical utility of gene-based prognostic signatures can be established. Two prospective clinical trials of prognostic gene signatures are now ongoing 1314 and, it is hoped, these will provide insight into the utility of widespread gene expression profiling as clinical tools.

A complementary approach uses a limited number of immunohistochemical markers as surrogates for the gene-based signatures. The 'intrinsic subtypes' 151617 have been reproduced repeatedly in this way and this methodology is likely to be a more realistic approach for routine diagnostic application 181920. In order to have clinical application, it is paramount that any marker shows good and poor prognostic groups independently of currently used standards or markers at stringent levels of statistical significance. With this in mind, we recently evaluated a panel of 13 biomarkers by immunohistochemistry (IHC) in a series of over 700 breast cancers and validated the findings in a second consecutive series of 983 cases on a tissue micro-array (TMA) 21. The panel included proteins whose genes were differentially expressed in the intrinsic subsets defined by microarray profiling 1516. We expected that such a panel would have greater prognostic power than any one individual marker but the results showed otherwise: only the anti-apoptotic protein BCL2 was required to predict overall survival independent of traditional parameters in both series. Furthermore BCL2 added to the prognostic power of the NPI.

The BCL2 protein is a member of the bcl family that regulate apoptosis. Its tumorigenic potential has been demonstrated in animal models 22 and is supported by the finding of over-expression of BCL2 in a variety of tumours and in lymphomas in which BCL2 acts as an oncogene 2324. In many solid organ tumours, including breast cancer, BCL2 paradoxically appears to exert a tumour suppressor effect, where its expression is associated with favourable prognostic features e.g. low grade, oestrogen receptor (ER)-positivity and good outcome 21. This has been a consistent finding in most reports but the association with outcome was limited to univariate analysis in many.

The aim of this study was to evaluate the prognostic role of BCL2 in breast cancer by sytematically reviewing the available evidence. To this end, we identified all published reports that assessed the relationship between BCL2 and outcome in breast cancer and performed a meta-analysis using standard statistical techniques.

Reports investigating the prognostic role of BCL2 in breast cancer were selected for review using a search of PubMed from 1994–2006 using the following criteria: '(bcl-2 and breast) AND (prognosis* [Title/Abstract] OR (first [Title/Abstract] AND episode [Title/Abstract]) OR cohort [Title/Abstract])' and also '(bcl-2 and breast) AND (incidence [MeSH:noexp] OR mortality [MeSH Terms] OR follow up studies [MeSH:noexp] OR prognos* [Title/Abstract] OR predict* [Title/Abstract] OR course* [Text Word])' limited to the English language. Three hundred and thirty three citations were retrieved. Eligible reports were those that examined the association between expression of BCL2 by IHC and either overall survival (OS) and/or disease free survival (DFS) in a clinical series of invasive breast cancer. Reports considered ineligible for the meta-analysis were reviews; those of in situ carcinoma or precursor lesions or of only male breast cancer; those that used cell lines or techniques other than IHC; and reports where either OS or DFS were not used as clinical endpoints; or where the association between another marker and outcome was being examined and data for BCL2 was not presented. Authors that published multiple reports on a single series were included once. The bibliography of the reports was also searched by hand.

Fifty-three eligible studies were identified. Small studies are more prone to publication bias and failure to report a hazard ratio (HR) can be taken as an indicator of study quality 25. We therefore included the seventeen studies that were based on at least 100 patients and had also reported an estimate of the HR 2126272829303132333435363738394041. The remaining 36 reports (Additional files 1 and 2) included 11 reports that were either based on a small number of cases (≤ 100) or were larger studies that did not report a HR but did provide sufficient data to estimate the HR according to the method of Parmer et. al. 42. These were analysed separately to determine if these studies would introduce significant bias 4344454647484950515253. The remaining 25 studies, including nine using small series, provided insufficient data for analyses and were excluded (Additional files 1 and 2).

Seventeen papers reporting 18 large case series with complete statistical data were included in the main analyses. These evaluated the role of BCL2 as a predictor of OS or DFS in 5,892 cases of breast cancer (Additional files 3 and 4). Median follow-up ranged from 0.2 to 472 months (average 92.1 months). In total, 2,619 cases of node-negative disease and 3,963 cases of node-positive disease were analysed. In 333 cases, nodal status was either unknown or not indicated. Five of the 18 series comprised consecutive cases unselected for specific characteristics 2129313339. The remainder consisted of cases that were either accrued onto clinical trials 21323537 or selected according to defined eligibility criteria i.e. node-negative 3041 or node-positive 262834363840 disease, or tumour size 27. Unfortunately, the number of studies was too small for formal meta-analysis for each of these end-points.

Fourteen studies reported on the effect of BCL2 on DFS of which three reported only unadjusted 323440 and six reported only multivariate adjusted 293033363739 hazards and five reported both 2627353841. Where multivariate analysis was performed the parameters included varied. Clinico-pathological variables were incorporated in most analyses with the following exceptions. Tumour size was excluded in two reports 3537. Nodal status was omitted in five studies 2630353741 although these were selected series that consisted entirely of either node positive 26, node negative 3041 or metastatic disease 3537. Tumour grade was also excluded in five reports 2627283738. The other biomarkers assessed varied with the exception of steroid hormone receptor status, which was incorporated in the majority of analyses (except 2729). p53 was included as a co-variable in 11 studies 2126272831323334373940 and HER2 in three 2126272831323334373940. A monoclonal antibody to assess BCL2, either from Dako (Clone 124) 2126272831323334373940, Novacastra 29 or Dakopatts 30363841 was used by most investigators. A polyclonal antibody was used in one study 35. Cytoplasmic staining was scored using a dichotomous scoring system in all studies with a cut-off for positive status between 10% and 40%. Berardo et al. 34 applied both continuous and dichotomous system and showed that the former was associated with independent significance but not the latter.

Eight studies including 2,285 patients reported the effect of BCL2 expression on DFS in analyses unadjusted for other risk factors (Figure 1, Table 1). The unadjusted HR estimates for DFS from these studies ranged from 0.85 – 3.03 and all but one of these was significant at a nominal p < 0.05. The pooled random effects estimate was 1.66 (95% CI = 1.25 – 2.22). However, there was evidence for significant heterogeneity amongst the studies (Q = 22.4, 7 degrees of freedom (df), p = 0.002), which might be expected given the difference in populations being studied and experimental methods used. Five studies 2634353841 included node-positive disease only accounting for 1,659 cases. Notwithstanding that, the heterogeneity was largely due to the study reported by Mottolese et al. 32, which was the only study to report a better outcome for BCL2 negative tumours. After the exclusion of the latter, the pooled estimate of HR was 1.74 (95%CI = 1.46–2.07) with no significant heterogeneity (Q = 6.6, 6 df, p = 0.36). There was no evidence for publication bias (p = 0.14). Eight studies were excluded from the primary analysis because they were small or because the HR was not reported but could be estimated from the data presented 4344454647485253. The pooled HR for these studies was 2.11 (95%CI = 1.62 – 2.77). This is significantly different from the estimate for the other studies (p = 0.02) suggesting, as predicted, a substantial bias in the HR estimates for the smaller studies and those with less completely reported data. If, however, all studies were analysed together the pooled HR estimate was 1.59 (95%CI = 1.41 – 1.81).

Eleven studies including 2,105 patients reported the effects of BCL2 on DFS adjusted for other prognostic factors 2627293033353637383941 (Table 1). The adjusted HR estimates were generally close to one but ranged from 1.10 to 3.26. In five of these studies incorporating 950 cases, BCL2 was an independent predictor of DFS. The pooled random effects HR estimate for these studies was 1.58 (95%CI = 1.29–1.94) without evidence for significant heterogeneity between the studies (Q = 17.2, 10 df,p = 0.07) or publication bias (p = 0.78). Four of the studies that failed to show an independent association between BCL2 and DFS used the Dakopatts or a polyclonal antibody.

The effect of BCL2 expression on OS was evaluated in 12 studies in 11 reports (Table 1). HRs were unadjusted in three of these 323440 and adjusted for other variables in four 28363739 and both unadjusted and adjusted HRs were given in four other reports 21273141. The eight studies from seven reports 21273141 incorporating 3,910 cases where the expression of BCL2 was unadjusted for other variables produced hazard estimates ranging from 0.99–4.31 (Figure 2), of which 6 were statistically significant with a pooled estimate of risk of 1.64 (95%CI = 1.36–1.97). There was evidence of statistical heterogeneity (Q = 17.5, 7 df, p = 0.015) that again was virtually entirely due to the contribution of the report by Mottolese et al. After exclusion of this study there was little change in the HR (pooled estimate 1.73, 95%CI = 1.48–2.02) and no evidence for heterogeneity (Q = 9.98, 6 df, p = 0.13) or publication bias (p = 0.87).

Six studies were excluded from the primary analysis because they were small or because the HR was not reported but could be estimated from the data presented 444849505152. The pooled HR for these studies was 3.42 (95%CI = 2.41 – 4.86). This is significantly different from the estimate for the other studies (p < 0.001), again suggesting a substantial bias in the HR estimates for the smaller studies. If, however, all studies were analysed together the pooled HR estimate was 1.99 (95%CI = 1.62 – 2.45).

Nine series comprising 3,624 cases were used for the meta-analysis of the expression of BCL2 on OS adjusted for other parameters 2127283136373941. BCL2 was an independent predictor of outcome in four of these 213639 totalling 2,190 cases. A narrow range of estimates was observed from 1.10 to 2.49 with a pooled estimate of 1.37 (95%CI = 1.19–1.58). There was borderline evidence of heterogeneity between the studies (Q = 15.48, 8 df, p = 0.05) and evidence of publication bias (p = 0.022).

The published data incorporating 5,892 cases strongly support the prognostic role of BCL2 as assessed by IHC in breast cancer showing that it is associated with both DFS and OS (pooled HR estimates of 1.66 and 1.64 respectively). These effects were slightly attenuated but still significant in multivariate analyses (adjusted HRs of 1.58 and 1.37 for DFS and OS respectively), showing that this effect is independent of lymph node status, tumour size and tumour grade as well as a range of other biological variables on multivariate analysis.

It is possible that the result of the meta-analysis could have been influenced by differences between the 18 studies. The finding of study heterogeneity in HR estimates is unsurprising and, even though the study by Mottolese et al. accounted for much of the heterogeneity for DFS, there were substantial differences between the other studies that could have influenced the result. Different cohorts of patients and treatment regimes were used. Five studies used unselected consecutive series 2129313339 and these patients were managed according to standard treatment protocols. In all of these reports, BCL2 was an independent predictor of either DFS 293339 or OS 213139 on multivariate analyses. Treatments varied where cases were part of a clinical trial 2126323537 although BCL2 was still an independent predictor of DFS and OS in the majority (2637 and 2137 respectively). The cause of the heterogeneity from the study of Mottolese et al. is unclear. The series was part of a clinical trial that examined the influence of a number of patho-biological factors on response to adjuvant therapy. Patient characteristics, type of antibody and scoring system were similar to those used by others 21. However, the authors showed an association between BCL2 positivity and adverse outcome, which is at variance with most other published reports and with all others included in the meta-analysis.

The types of antibodies employed and cut-offs used to define immunohistochemical positivity also varied between reports. The effect of each of these is difficult to assess but could have contributed to the observed heterogeneity. Clone 124 was the most commonly used and was used in the majority of studies where BCL2 was an independent predictor of DFS or OS 212627283133373940. In contrast, most reports that used the Dakopatts monoclonal antibody or a polyclonal antibody 30353641 failed to show a statistically significant association with DFS or OS, although the direction of effect was the same. There was little difference in the scoring systems between the studies although the different cut-off points varied from 10% to 40%. The inclusion of different combinations of variables in multivariate analyses could also have affected the results of a pooled estimate of the adjusted HRs. The purpose of the meta-analysis was to establish if the association of BCL2 with prognosis is independent of other confounders. The majority of studies included the most important variables in their analysis, which suggests that the pooled adjusted estimate of hazard is robust and supports BCL2 as an independent risk factor. In addition, the similarity between the HRs observed for BCL2 in both univariate and multivariate analyses, where both types of analyses were performed also attest to its prognostic role.

The effect of bias on the meta-analysis should also be considered. Seventeen eligible reports provided data suitable for the meta-analysis. Thirty six eligible studies were excluded, including studies using small series and those that did not provide minimum data for the pooled analysis. The exclusion of small studies may have minimised the effect of publication bias – the non-publication of studies with null results – by not including reports of small series that are more likely to be published if they show a positive result. This is supported by our findings that the HRs reported by smaller studies were systematically larger than those reported by the larger studies. The inclusion of only large studies and those that meet minimum quality criteria in the meta-analysis maximised the chance of the pooled estimate of HR representing the true HR. Whether we have been able to retrieve all relevant reports in the literature is unclear. In particular, it is possible that our search strategy will have missed studies evaluating BCL2 as one of multiple markers, but where no association with outcome was detected. This is perhaps most likely to occur in studies using TMAs. However, if large negative studies had been missed, a substantial degree of publication bias (perhaps a misnomer here as the data may be published and not retrieved) would have been detected, whereas we found little evidence for this. It seems unlikely that missing data bias either through non-publication, or through failure to identify and retrieve the data, would have led to a substantial overestimation of the meta-analysis HR estimates.

The mechanisms through which BCL2 might exert its protective effect in breast cancer are unclear. Whether it is consequent upon its role in apoptosis or whether non-apoptotic functions are involved is unknown. The anti-apoptotic role of BCL2 is well characterised but its function in cell cycle control has received less attention. The latter is well supported by cell line studies that show BCL2 expression delays G1 progression and G1-S transition by prolonging G0 and is capable of growth inhibitory effects analogous to those of p53 2457. It is postulated that the dominance of one of these functions over another may depend on the cell type and physiology and that the anti-proliferative effect translates into a tumour suppressor role in solid epithelial tumours including breast cancer. Furthermore, it remains to be established if other proteins e.g. Wnt11 58 or bcl homologues potentiate the tumour suppressor role of BCL2 in breast cancer. At least 20 bcl proteins are known that may act in a synergistic manner, analogous to the co-operative effect of MYC and BCL2 in lymphomagenesis.

The therapeutic implication of our findings and how BCL2 might improve prognostication and/or selection of patients for treatment remain to be determined. The data presented here shows that BCL2 has the potential to improve patient stratification and guide patient management. Currently, patients are stratified into different treatment categories based primarily on nodal status, tumour size, tumour grade, receptor status, patient age and, to a lesser extent, tumour type. A range of guidelines e.g. St. Gallen, National Institutes of Health Consensus Development Panel, National Comprehensive Cancer Network (NCCN) Clinical Practice Guidelines and prognostic systems e.g. NPI, ADJUVANT! are used internationally, each of which combines these features in different ways to assign patients to risk categories. Recommendations for systemic treatment are made by some, for example, the NCCN recommends adjuvant chemotherapy for tumours > 1 cm in combination, with adjuvant hormonal and traztuzumab depending on receptor status, but not for small (≤ 5 mm), node-negative invasive tumours or for node-negative grade 1 tumours between 6 mm and 1 cm. In the UK, the NPI is used widely to inform decisions on adjuvant therapy 5960, however, there is no agreement on cut-off values. The prognostic value of the NPI is supported from validation studies but no new prognostic factors have been shown that add substantially to its use. Taken in conjunction with previous data, our results suggest that BCL2 could improve the stratification of patients by the NPI and could separate both the moderate and the poor prognostic groups each into two prognostic categories. We showed previously that the prognostic effect of BCL2 status is maximal in the first five years after a diagnosis of breast cancer and wanes thereafter, suggesting that its utility as a diagnostic adjunct may be limited to this period. Even if BCL2 were to be used in clinical decision making, the most appropriate cut-off used to assign positivity is uncertain. Our previous report 21 indicated that a dichotomous scoring system using a 10% cut-off provided the most parsimonious fit for predicting overall survival using a cox regression analysis, but published data are insufficient to replicate this.

The number of studies included in the meta-analysis with different treatment endpoints was too small to perform a meaningful analysis of its predictive role. However, there is an emerging consensus that BCL2 plays a key role in determining response to endocrine therapy and chemotherapy 61. BCL2 is an oestrogen responsive gene 5862 and many clinical studies have shown an association with favourable response to endocrine therapy 406163. BCL2 is a component of the 21-gene signature used to predict recurrence in tamoxifen-treated ER-positive node-negative breast cancer and its role is being evaluated prospectively in the TAILORx trial 14.

In summary, we have shown that BCL2 is an independent prognostic marker in two large series of breast cancer 21 and, here, provide an estimate of the average size of this association with outcome in a meta-analysis on 16 other series totalling over 5,000 patients. Using this approach, we have demonstrated the prognostic power of a single marker in breast cancer. Large studies, both observational cohorts and clinical trials, are now urgently needed to test whether BCL2 and multiple other markers can provide prognostic information in addition to currently used standards and also to establish if BCL2 has clinical utility.

CI: confidence interval; df: degrees of freedom; DFS: disease free survival; EC: epirubicin and cyclophosphamide; G-CSF: granulocyte colony stimulating factor; HER2: Human epidermal growth factor receptor; HR: hazard ratio; IHC: immunohistochemistry; NCCN: National Comprehensive Cancer Network; NPI: Nottingham Prognostic Index; OS: overall survival; TAILORx: Trial Assigning Individualized Options for Treatment (Rx); TMA: tissue micro-array; TNM: tumour, nodal status, metastasis; UICC: Union Internationale Contre le Cancer.

The authors declare that they have no competing interests.

MJW performed the search of pubmed and organised the data. GMC organised the data, drafted and finalised the manuscript. PDPP performed the meta-analysis. PDPP and CC both contributed to writing and finalising of the manuscript. All authors read and approved the final manuscript.