Clinical and histopathological factors such as lymph node status, tumor size, histological grade, age, and expression of estrogen receptor (ER) and Her2 have traditionally guided treatment decisions of patients with operable breast cancer 1 2 . Various prognostic models are based on these factors, for example the Nottingham Prognostic Index (NPI) 3 4 , Adjuvant!Online 5 6 and others 7 . Despite providing excellent estimates of the average risk of recurrence, there remains substantial variation in outcome which may be explained by molecular differences among these tumors 8 9 .

DNA-chip based expression analyses have confirmed the heterogeneity of breast cancer and allowed the development of clinically relevant gene "signatures" or "profiles" 10 11 12 13 14 15 16 17 18 19 20 . Such profiles are being implemented widely in routine patient care even though many signatures were developed and validated on heterogeneous patient cohorts with respect to stage of disease and therapy. The utility of gene signatures as part of the decision making process is being validated in ongoing studies (TAILORx 21 and MINDACT 22 ). Most profiling studies are based on fresh-frozen (FF) or RNAlater conserved tissue. Such material must be collected and processed separately after surgery, complicating the implementation of molecular analyses into the clinical workflow. Procedures based on formalin-fixed, paraffin-embedded (FFPE) material simplify the acquisition of tumor material and can easily be established as part of the routine pathological procedures. In addition, FFPE tissues collected in the framework of clinical trials could be a valuable resource for future research.

We prospectively selected genes from publicly available microarray data and developed molecular scores representing the ER, progesterone receptor (PgR), Her2 and proliferation (PRO) status, and the overall risk of recurrence (RISK). The reproducibility and robustness of the molecular scores was validated by comparing expression data with RNA from FF and FFPE material of 82 tumors. Molecular scores were determined from 342 ER positive tumor samples of the BIG 1-98 clinical trial. Multivariate Cox proportional hazard models revealed that molecular scores are independent prognostic factors to estimate disease free survival (DFS).

To assess the quality of expression profiling from FFPE material, matched FF and FFPE samples from 82 human breast cancers were used. Histopathological information was irreversibly anonymized according to Swiss law. Independent FFPE blocks and corresponding clinical data of 437 Swiss participants of the trial BIG 1-98 were provided by the International Breast Cancer Study Group. The ethics committees and required health authorities of each participating institution approved the study protocol, and all patients gave written informed consent (ClinicalTrials.gov number, NCT00004205) 23 . Retrospective tissue collection was carried out in accordance with institutional guidelines and national laws. The patient and tumor characteristics of these patients were similar to the entire BIG 1-98 population (Table 1). BIG 1-98 is a randomized controlled clinical trial of adjuvant hormonal therapy for postmenopausal patients with endocrine-responsive breast cancer comparing 4 arms: 5 years of tamoxifen, 5 years of letrozole, two years of tamoxifen followed by 3 years of letrozole, or vice versa 24 25 26 . All the patients from the BIG 1-98 were treated by mastectomy or breast conserving surgery 24 25 26 . The available paraffin blocks contained material derived from representative tumor regions.

Gene expression profilings define clinically relevant gene signatures 15 17 37 38 . For the present work, we selected genes correlating with the ER, PgR, Her2 and proliferative status using a meta-analysis of gene expression profiles 28 . The prognostic power of resulting gene expression scores for ER, PgR, proliferation and overall risk of recurrence was validated using tissues and clinical data from a representative subset of participants of trial BIG 1-98 confirming the correlation structure of these genes and their association with clinical and outcome variables.

Multiple genes representing each score were quantified by qRT-PCR. RNA from 82 matched FF and FFPE tissues were compared by qRT-PCR on TLDAs. The mean increase of raw Ct values between RNA from FF and FFPE tissues was 1.3 units. This is similar to the findings of Cronin and co-workers (+2.0 units) in a comparable setting 39 . Duration of formalin fixation, storage time and conditions influence the quality of RNA derived of FFPE tissues with direct effects on the sensitivity of subsequent PCR reactions 40 . However, normalization effectively compensated for this shift of Ct values (Additional File 2, Fig S1 and Additional File 3, Figure S2).

The mean expression of eight genes related to ER and five genes related to PgR were used to calculate the ER_8 and PGR_5 scores. Scores representing different functional categories were combined in RISK_25 score. The molecular scores determined from 82 paired samples of FF and FFPE tumors were highly concordant, as were molecular scores and immunohistochemically assessed parameters demonstrating the reliability of the procedure.

Molecular scores were validated in an independent set of tumor tissues from 342 participants of trial BIG 1-98. In contrast to histological analyses which can also be performed from tissue sections that contain considerable normal, stromal or fat components the architecture of the tissue is completely lost during work up for molecular analyses and therefore, it was important to exclude samples with inadequate tumor content. A histological section was taken from the immediate vicinity of each sample that was used for molecular analyses. Each section was assessed by an experienced pathologist (H.J.A.) and molecular analyses were restricted to samples containing at least 30% tumor cells. For comparison, RNA was also isolated from tumor-surrounding cells which led to rather poor RNA recoveries from comparable tissue areas (data not shown). However, this does not exclude that tumor-surrounding cells may have a limited impact on molecular scores in such analyses. Contamination by non-tumor cells may be reduced by macrodissecting tumors before RNA isolation and molecular assessment. The same procedure would also make tumors accessible to molecular analysis when sections contain less than 30% tumor cells.

Classification of patients by low and high PRO_10 and RISK_25 scores corresponded to low and high risk of recurrence. PRO, RISK and PGR scores were prognostic for DFS not only in the entire patient population but also in a subpopulation of patients with node positive disease (Figure 3D and 3E). We provide evidence independent of Genomic Health™ that a RISK score based on similar biological processes as the recurrence score (RS), but with other genes selected through a different procedure, can predict DFS 29 41 42 . In contrast to the RS which was validated with tamoxifen-treated patients, PRO_10, RISK_25 and PGR_5 scores were validated with patients treated with tamoxifen, letrozole or a sequence of both drugs; therefore, they may apply to patients who received either of these drugs.

Histological grading is an important factor in estimating the risk of recurrence of patients with breast cancer 2 43 . Recently, Sortiriou and colleagues have developed the gene expression grade index (GGI) based on the expression of 97 genes related to proliferation. They demonstrated that grade II cancers are comprised of tumors which are similar to genomic grade I or grade III with corresponding clinical outcomes 16 44 . Our findings agree with these observations as grade II tumors could be further classified into low and high risk of recurrence by 10 genes (PRO_10) (Figure 3C) or even by 5 genes (PRO_5 score) (data not shown). Seven of the PRO_10 and three of the PRO_5 genes are also part of GGI. The PRO_5 genes (Table 1) corresponded to the proliferation-related genes of the RS 29 . The assessment of gene signatures related to proliferation such as GGI or PRO scores is of special interest in ER positive, grade II breast cancer for whom therapeutic decisions are often difficult. Both, GGI and RS were shown to be associated with response to chemotherapy 45 46 . In contrast to GGI which requires FF tumor material, PRO scores or RS can be determined from a few microtome slices or cores such as used for tissue microarrays 47 . Material for molecular analysis can be taken from the same FFPE tissue block used for histological and immunohistochemical analyses without interfering with clinicopathological workflow.

The prognostic value of Ki-67 in early breast cancer was recently confirmed 48 . However, Ki-67 is not used uniformly in clinical practice 49 50 as it appears to be difficult to agree on cut-off values separating high and low proliferation tumors or on its value in assisting the choice of adjuvant therapy 50 51 . Therefore, instead of dichotomizing Ki-67 it may be more feasible to use Ki-67 as continuous variable 52 . Here, we made a comparison between centrally assessed Ki-67 LI and a qRT-PCR based proliferation signature. The PRO_10 score correlated with Ki-67 LI, and both were significant predictors of DFS in univariate Cox analyses. In multivariate models however, PRO_10 offered superior prognostic value and outperformed Ki-67 LI (Table 3). Moreover, the PRO_10 score added independent prognostic information to anatomical staging.

PgR, as measured by immunohistochemistry 30 or microarray analysis 53 , was shown to positively correlate with prognosis. Here we show that the molecular PGR_5 score was also positively associated with DFS (Figure 4) and added independent prognostic information to anatomical staging and PRO_10 (Table 3). Thus, PGR_5 and PRO_10 scores independently predict prognosis in the BIG 1-98 population.

Compared to immunohistochemically assessed parameters, qRT-PCR based scores are quantitative, relatively independent on operator expertise and less affected by inter-observer variability. The procedure is simple, economical and can be standardized easily with good control genes, reference samples and quality control procedures.

The results of this study are based on a limited number of patients and follow-up time (60 months). Similar analyses with independent, larger sample sizes and more mature follow-up data are planned to further consolidate the prognostic and possibly predictive value of the proposed scores in each treatment arm separately.

Gene expression profiling has improved the understanding of molecular subtypes of breast cancer. FFPE material is not widely used although it may facilitate and speed up the development and validation of novel gene signatures due to the availability of well-characterized tissues from numerous clinical trials 54 55 . The same material can be used for molecular diagnostics. The investigation of gene signatures may become more important in the future as an increasing proportion of agents under development for breast cancer treatment have defined molecular targets. Early integration of biomarker analysis in the drug development process has the potential to improve the specificity and efficiency of novel therapeutics. This opens the possibility to further individualize therapy of patients with breast cancer.

We define four molecular scores based on quantitative measurement of gene expression with RNA derived of FFPE tissues. The genes for each score were selected from a large meta-analysis of microarrays. The genes do not coincide with genes used for other molecular scores like the RS (except genes that were previously used as immunohistochemical markers such as ER, PgR or Her2). Two of the described scores are shown to be independent predictors of disease-free survival of postmenopausal patients with operable, estrogen receptor positive breast cancer. The proliferation-associated score outperforms the Ki-67 labeling index measured by immunohistochemistry.

AUC: area under the (ROC) curve; CI: confidence interval; DFS: disease-free survival; ER: estrogen receptor; FF: fresh frozen; FFPE: formalin-fixed, paraffin embedded; HR: hazard ratio; IHC: immunohistochemistry; GGI: gene expression grade index; LI: labeling index; LRT: likelihood ratio tests; PCR: polymerase chain reaction; RIN: RNA integrity number; PgR: progesterone receptor; ROC: receiver operating characteristic; RS: recurrence score; TLDA: TaqMan Low Density Arrays.

JA, VP, MD, PW, AB, AO, AGH, GV, HJA, SA and RJ declare that they have no competing interest. B.T. holds stocks from Novartis (Ciba Geigy) since 1990.

JA, SA and RJ organized the study, planned the experiments and wrote the manuscript. SA and BT organized samples from the International Breast Cancer Study Group. AO and AB carried out RNA isolations, quality controls and gene expression measurements. VP, PW, MD and AGH carried out the statistical analyses. HJA and GV were responsible for histological assessment of stained sections. All authors contributed to the manuscript, they read and approved the final manuscript.