We included all breast cancer patients that were operated on at the Karolinska Hospital from 1 January 1994 to 31 December 1996 (n = 524), identified from the population-based Stockholm–Gotland breast cancer registry established in 1976. Available tumor material was frozen on dry ice or in liquid nitrogen and was stored in -70°C freezers. Figure 1 shows the details of various exclusions leading to the final 159 patients for analysis. The ethical committee at the Karolinska Hospital approved this microarray expression project.

The different reasons for exclusion were not influenced by age at diagnosis (Table 1). The 231 tumors that were not analyzed using expression profiling had a lower mean diameter, had fewer mean affected lymph nodes, and had fewer individuals with recurrent disease at the end of the study period (Table 1). For those excluded for other reasons, there did not seem to be a selection based on age or stage of the disease, compared with those patients included in the study (Table 1).

The Stockholm–Gotland Breast Cancer Registry, supplemented with patient records, were examined for information on the tumor size, the number of retrieved and metastatic axillary lymph nodes, the hormonal receptor status, distant metastases, the site and date of relapse, initial therapy, therapy for possible recurrences, the date and cause of death. Tumor sections from the primary tumors from patients with array profiles were classified using Elston–Ellis grading 18 by a blinded pathologist (HN).

In the adjuvant setting tamoxifen and/or goserelin is normally used for hormonal treatment, but mostly intravenous cyclophosphamide, methotrexate and 5-fluorouracil (CMF) on days 1 and 8 was used as adjuvant chemotherapy, except in high-risk patients who were offered inclusion in the Scandinavian Breast Group 9401 study 19. After primary therapy, patients were recommended to have regular clinical examinations and yearly mammograms, in addition to laboratory and X-ray tests guided by clinical signs and symptoms. Patients were normally followed for 5 years. Patients followed up outside the Karolinska Hospital were tracked using a unique personal identification number. There was no loss to follow-up.

The relapse site, date of relapse, relapse therapy and date of death were ascertained in May 2002. The average follow-up was 6.1 years. Cause of death was coded as death due to breast cancer (including those with distant metastases but dying from other causes), death due to other malignancies and death due to nonmalignant disorders. Through the population-based Swedish Cancer Registry, second primary malignancies were identified.

For validation we used population-derived primary breast cancer patients receiving primary therapy from 1987 to 1989 in the county of Uppsala, Sweden 202122. From the initial set of 315 patients, representing 65% of all breast cancer patients in Uppsala county during these years, we were able to obtain quality-controlled RNA expression profiles from 260 frozen tumors (including two patients with neoadjuvant tamoxifen) (Fig. 1). A further follow-up of events was carried out with a 1999 deadline. Seventy-six lymph-node-positive patients received adjuvant, mostly intravenous, 3-weekly CMF-based therapy (premenopausal patients) or adjuvant tamoxifen (postmenopausal patients) 20. Some node-negative patients also received adjuvant therapy as previously described 20. One hundred and thirty-five node-negative patients did not receive adjuvant therapy. All tumors have been confirmed to have invasive cancer and have been graded according to Elston–Ellis, except for one patient with missing primary tumor slides but with the presence of axillary lymph nodes, thus confirming invasiveness. The ethical committee at the Karolinska Institutet approved this RNA expression study.

RNA extraction was performed according to the RNeasy mini protocol (Qiagen, Hilden, Germany). In brief, a portion of the deep frozen tumor was cut into minute pieces and transferred into test tubes (maximum 40 mg/tube) with RLT buffer (RNeasy lysis Buffer, Qiagen, Hilden, Germany), followed by homogenization for around 30–40 s. Proteinase K was then added and the samples were treated for 10 min at 55°C. This step was introduced during the project 23 because most initial preparations without this step resulted in either poor RNA yield and or poor RNA quality. Total RNA was then isolated using Qiagen's microspin technology. DNase was also added to some samples to further increase the RNA quality. The quality of the RNA was assessed by measuring the 28S:18S ribosomal RNA ratio using an Agilent 2100 bioanalyzer (Agilent Technologies, Rockville, MD, USA). All samples with RNA of high quality were then stored at -70°C until microarray analyses.

Preparation of in vitro transcription products and oligonucleotide array hybridization and scanning were performed according to the protocol of Affymetrix (Santa Clara, CA, USA). In brief, the amount of starting total RNA for each probe preparation varied between 2 and 5 μg. The in vitro transcription reactions were performed in batches to generate biotinylated cRNA targets, which were subsequently chemically fragmented at 95°C for 35 min. Fragmented and biotinylated cRNA (10 μg) was hybridized at 45°C for 16 hours to Affymetrix high-density oligonucleotide array human HG-U133 set chips. The arrays were washed, and were then stained with streptavidin–phycoerythrin (final concentration, 10 μg/ml). The array was then scanned according to the manufacturer's instructions (Affymetrix Genechip^® Technical Manual, 2001; Affymetrix). The scanned images were inspected for the presence of obvious defects (artifacts or scratches) on the array. In the case of visible microarray artifacts, the sample was rehybridized and rescanned on new chips using the same fragmented probe. The raw expression data were normalized using the global mean method 24.

A statistical data filter was applied to reduce noise and to obtain a useful and relevant probe set to identify markers that were highly correlated to clinical parameters. The detail is provided in the Supplementary Report (Additional file 1). This led to 6573 final probe sets for analysis, consisting of 3393 probe sets from U133A and 3180 probe sets from U133B. All analyses were performed using natural-log-transformed expression values.

The primary statistical analysis was based on comparing good prognosis and poor prognosis, where poor prognosis was defined as distant relapse or death from any cause within 5 years of diagnosis. For comparison, a secondary analysis was performed limiting poor prognosis to distant relapse and death due to breast cancer. The secondary classification resulted in seven patients switching from the poor to good prognosis group.

In order to maximize the statistical power, we initially used the whole Stockholm cohort (n = 159) as the training set for choosing an optimal gene set and identifying risk groups in the hierarchical cluster analysis. This cohort was a mixture of patients without adjuvant therapy and 126 patients treated with CMF, tamoxifen, megesterolacetat, goserelin or some combinations thereof. We checked the consistency of the analysis of the whole cohort against the subset of patients treated with tamoxifen and its combinations (n = 104) and against all systematically untreated patients (n = 33).

To validate the results we used independent datasets from Uppsala, consisting of 76 node-positive adjuvant-treated patients and 135 node-negative untreated patients, and from the Dutch study of 78 untreated node-negative patients 16, referred to as the van't Veer data.

An optimal set of predictors was chosen using a leave-one-out cross-validation procedure performed on the training set (Additional file 1). Class prediction using k genes was carried out using a diagonal linear discriminant analysis method 25, which is a variant of the standard maximum-likelihood discrimination rule. The class predictor score S is computed from the top k genes. A patient with S > 0 is assigned to the poor-prognosis group, and otherwise to the good prognosis group. We will thus refer to S as the poor-prognostic score or the risk score.

To investigate whether the risk score had an independent predictive value over the standard clinical variables, the risk score S (high–low, with 'high' defined as S > 0) was included in a multivariate logistic regression analysis with 5-year status as the outcome variable. To obtain unbiased estimates, the scores for patients in the training set were computed from the leave-one-out procedure; because of dependence between samples, however, this procedure tends to produce optimistic standard errors 26. We did not attempt to correct the standard errors, because the result was also validated in independent datasets. The clinical variables were the age at diagnosis, tumor grade, tumor size and lymph node metastasis, estrogen receptor status (positive–negative) and progesterone receptor status (positive–negative). The tumor size and lymph node metastases were entered into the model in the form of a stage variable. These clinical predictors were initially compared between the good-prognosis and poor-prognosis groups.

Unsupervised hierarchical clustering of the training data was used to identify flexible risk groups; here we used the Euclidean distance with complete linkage. For validation data, we used supervised clustering based on the assignment of samples to the cluster with the closest centroid. The standard Euclidean distance was used for Uppsala datasets, but for the van't Veer dataset, because of different scales and possible outliers, the distance was based on Spearman rank correlation.

To obtain a better description of the prognosis of the patients during the follow-up, we also performed survival analysis, enabling us to use full survival information not just the 5-year status. The Cox proportional hazard model was used to assess the additional contribution of the prognosis score after adjusting for the clinical variables.

Our analytic approach improved the previous studies of breast cancer prognosis using microarray gene expression data by jointly analyzing adjuvant-treated patients and untreated patients. A previous Dutch study identified 70 prognostic genes from an analysis of untreated patients 16, but provided no indication of who might fail to respond to adjuvant therapy. Other studies involving treated patients also did not discuss the treatment assessment 27.

Any evaluation of adjuvant therapy must consider three types of patients: type A, those patients who do well without treatment; type B, those patients who do poorly without treatment, but may benefit from treatment; and type C, those patients who do poorly despite treatment. It is clear that type-A patients should not be treated, while type-C patients require new treatment protocols that were not available during the study periods 1987–1989 and 1994–1996, respectively. Our results indicate that the low-risk cluster consists mostly of type-A patients and the high-risk cluster consists mostly of type-C patients. The medium-risk group does not provide such clear information. In the present study we identified that around 30% of the patients in the Stockholm cohort were of low risk, hence requiring no adjuvant therapy, and that almost 40% were of high risk, for whom the existing adjuvant therapy failed. Almost three-quarters of the patients therefore did not benefit from treatment. To our knowledge, no previous study has stated this as explicitly.

One weakness of the current study as well as all of the previous studies is the inability of the gene expression profile to identify patients that will actually benefit from treatment (type-B patients). Without a randomized trial, this goal appears difficult to reach.

Twenty-two of the 64 genes have unknown function, while the other genes represent various biological pathways such as DNA replication and transcription, cell-cycle regulation, cell adhesion and metastasis (Table 5). Among the genes associated with low-risk tumors we found CDKN1C, PDGFD, IGF1, HOXA5, SLIT2 and PTN, and many of the genes associated with high-risk tumors appeared to be involved in cell-cycle regulation (including CDC2, CDC20, BUB1B, PRC1 and RRM2) and in transcription (such as TOP2A).