Supporting materials and methods


Patients and tissue samples

The majority of the tumor samples analyzed in this study (Norway/Stanford cohort) was sampled in two independent prospective studies on response to chemotherapy of locally advanced breast cancer (T3/T4 and/or N2). As part of the design of these studies, two independent biopsies were obtained from the majority of the tumors, one prior to therapy and the second from the patients who had surgery after their chemotherapy. Altogether, paired biopsies were available from 54 patients, however; only one tumor sample from each patient was used for the classification. From the first cohort of patients treated with doxorubicin monotherapy (1), 55 tumor samples were analyzed, of which 51 have been previously published (2). From the second cohort of patients treated with 5-fluorouracil and mitomycin C (samples labeled with FU) (3), 32 tumor samples were analyzed. The remaining 28 samples, most of which were previously reported, were primary tumor specimens collected either at Stanford or in Norway. See supplemental table 1 for detailed information on all patients/samples. Altogether, 122 tissues samples were included in the analysis, of which 77 carcinomas and 7 non-malignant tissues have previously been described (2).

Microarray analysis

Preparation and purification of RNA, labeling and array hybridization was performed as described (2). All samples not previously published were analyzed using cDNA arrays containing 42,000 features, manufactured at the Stanford Microarray Core Facility ( The primary data for all experiments are stored in the Stanford Microarray Database  ( (4), and can be accessed at ( See supplemental table 1 for details on array and reference batches.

Selection of "intrinsic" genes

As a basis for classification, we wished to identify genes whose expression levels varied significantly from tumor to tumor, but which varied relatively little between successive samples of the same patient's tumor (in this case, two independent samples obtained from a tumor separated by 15 weeks of chemotherapy treatment) (5). We have used the term "intrinsic" (to the tumor) to refer to genes selected this way. Before "intrinsic" gene selection, the raw data were filtered to include only features with signal intensity 1.5 over background in both Cy5 and Cy3 channels, and for which for which this signal intensity criteria were met in at least 80% of the samples. In addition, the data file was adjusted for array batch differences as follows; on a gene-by-gene basis, we computed the mean of the non-missing expression values separately in each batch. Then for each sample and each gene, we subtracted it's batch mean for that gene. Hence, the adjusted array would have zero row-means within each batch. This ensures that any variance in a gene is not a result of a batch effect. Using 45 tumor pairs (including two primary-lymph node pairs) and the approximately 8000 genes common for all experiments, we computed for each gene the average "within-pair variance" (the average square before/after difference), as well as the "between-subject variance" (the variance of the pair averages across subjects). We then computed the ratio D=(within-pair variance)/(between-subject variance) and declared those genes with a small value of D to be intrinsic. The choice of a value of D as a cutoff was somewhat arbitrarily set at one standard deviation below the average, yielding a set 561 clones representing 540 genes (matched to UniGene by using SOURCE and includes 34 clones that matched to two separate UniGene clusters).

Data and gene filtering

Data (normalized log2 ratios) were obtained for all 122 tissue samples for the unflagged features using the identified "intrinsic" gene set. This reduced the list to 534 genes represented by 552 clones to be used for further data analysis. Rows (genes) were median-centered and both genes and experiments were clustered using an average hierarchical clustering algorithm (6). Replicate clones representing one single UniGene cluster were weighted before clustering. The cluster diagrams were visualized using TreeView (

Description of the two independent published data sets

van't Veer et al. (7), analyzed tumor samples from 117 young breast cancer patients, of whom 18 carried a BRCA1 mutation and two carried a BRCA2 mutation. The sporadic patients had invasive breast tumors less than 5 cm (T1 or T2), no axillary metastases (N0) and were diagnosed before the age of 55 years. All patients had surgical treatment followed by radiotherapy. Five patients received adjuvant systemic therapy (three chemotherapy and two hormone therapy). Follow-up time was at least 5 years. A two-color hybridization protocol was used in which the reference sample was a pool of equal amounts of cRNA from each individual tumor sample. Samples were hybridized using a dye reversal technique to oligonucleotide microarrays (60-mers) representing approximately 25,000 human genes. By using a supervised classification, the authors identified a poor prognosis gene expression signature consisting of 231 genes, of which 70 were optimal in predicting a short time interval to distant metastases (within 5 years).

West et al. (8) analyzed 49 invasive ductal carcinomas between 1.5 and 5 cm in maximal dimension, most of which were stage 2. Tumors were identified either as positive for both estrogen and progesterone receptors or negative for both, and as positive or negative for tumor cells in the diagnostic axillary lymph node. No additional clinical information was available for these patients. The samples were analyzed on Affymetrix Human Gene FL genechip DNA arrays (25-mers). The authors used binary regression models combined with singular value decomposition to identify the top 100 genes in their data sets that maximally correlated with estrogen receptor status in one analysis and with lymph node status in another analysis.

Extraction of data from West et al.

From the total of 7129 genes measured in this study using Affymetrix oligonucleotide arrays, UniGene cluster IDs were extracted for 5023 genes by using SOURCE ( (9). Data from each array were scaled such that each had the same overall intensity (normalized to the median experiment in the data set). For each array, intensity measurements below 20 units were replaced with the value of 20, thus removing negative absolute intensities. Each absolute expression value in a given sample was converted to a ratio by dividing by its average expression value across all samples. Only those genes for which at least 10% of samples provided adequate data (based on the Absent/Present calls provided by the authors) were used for further analysis. Genes expressed at detectable levels in fewer than 10% of the samples were excluded. This reduced the number of genes represented in the data set to 3074, including 242 members of the "intrinsic" gene set

Computation of expression centroids

There were 500 genes in the intrinsic gene set that matched to a unique, single UniGene cluster by SOURCE. Centroids of the tumors' patterns of expression of these genes were computed for each of the 5 classes in the Norway/Stanford data set. These are profiles, consisting of the average gene expression for each of the 500 genes, in each of the 5 subtype classes. Only those tumors that showed the highest correlation with each other within a subtype were used for the calculation (supplemental table 3). We then computed the correlation of each sample from the two additional data sets published by van't Veer et al. (7), and West et al. (8), to each of these 5 centroids. In figures 2, 3 and 4, we colored the dendrogram branch representing each tumor according to the class of the centroid with maximal correlation.

Prediction of tumor subtypes using Norway/Stanford data as training set

Class prediction was performed using Prediction Analysis of Microarrays (PAM), which is a variant of nearest-centroid classification with an automated gene selection step integrated into the algorithm (10). PAM uses a parameter, D, to minimize the gene list used during prediction. During ten-fold cross-validation we iteratively increased D so that a value could be chosen for the final model. We chose the value of D that balanced prediction accuracy with a minimal set of genes. This value of D was used for training on the entire Norway/Stanford set (115 tumors) and predicting the class of each sample in the van t Veer et al. and West et al. data sets. Supplemental table 2 shows the results of using both a five-class (luminal A, luminal B, basal, ERBB2+ and normal breast) and a four-class (luminal A and B combined, basal, ERBB2+ and normal breast) predictor.





     1.    Geisler, S., Lonning, P. E., Aas, T., Johnsen, H., Fluge, O., Haugen, D. F., Lillehaug, J. R., Akslen, L. A. & Borresen-Dale, A. L. (2001) Cancer Res. 61, 2505-2512.

     2.    Sorlie, T., Perou, C. M., Tibshirani, R., Aas, T., Geisler, S., Johnsen, H., Hastie, T., Eisen, M. B., van de, R. M., Jeffrey, S. S. et al. (2001) Proc Natl Acad Sci U S A 98, 10869-10874.

     3.    Geisler, S., Borresen-Dale, A. L., Johnsen, H., Aas, T., Geisler, J., Akslen, L. A., Anker, G. & Lonning, P. E. (2003) Submitted to Clin Cancer Res.

     4.    Gollub, J., Ball, C. A., Binkley, G., Demeter, J., Finkelstein, D. B., Hebert, J. M., Hernandez-Boussard, T., Jin, H., Kaloper, M., Matese, J. C. et al. (2003) Nucleic Acids Res. 31, 94-96.

     5.    Perou, C. M., Sorlie, T., Eisen, M. B., van de, R. M., Jeffrey, S. S., Rees, C. A., Pollack, J. R., Ross, D. T., Johnsen, H., Akslen, L. A. et al. (2000) Nature 406, 747-752.

     6.    Eisen, M. B., Spellman, P. T., Brown, P. O. & Botstein, D. (1998) Proc Natl Acad Sci U S A 95, 14863-14868.

     7.    van't Veer, L. J., Dai, H., van de Vijver, M. J., He, Y. D., Hart, A. A., Mao, M., Peterse, H. L., van der, K. K., Marton, M. J., Witteveen, A. T. et al. (2002) Nature 415, 530-536.

     8.    West, M., Blanchette, C., Dressman, H., Huang, E., Ishida, S., Spang, R., Zuzan, H., Olson, J. A., Jr., Marks, J. R. & Nevins, J. R. (2001) Proc Natl Acad Sci U S A 98, 11462-11467.

     9.    Diehn, M., Sherlock, G., Binkley, G., Jin, H., Matese, J. C., Hernandez-Boussard, T., Rees, C. A., Cherry, J. M., Botstein, D., Brown, P. O. et al. (2003) Nucleic Acids Res. 31, 219-223.

   10.    Tibshirani, R., Hastie, T., Narasimhan, B. & Chu, G. (2002) Proc Natl Acad Sci U S A 99, 6567-6572.