To perform a comprehensive evaluation of the mini-array results, we compared in Figure 3 the current classification to the good and poor prognostic profiles with that of the originally published classification for each sample. Results from the original study are shown (X-axis) plotted against those obtained from the customized mini-array (Y axis) 6. The data generated using the diagnostic test is highly similar (Pearson correlation of 0.92, p < 0.0001) to the original published data. The overall accuracy of the diagnostic test was determined by calculating the odds ratio for the development of distant metastases within five years. The odds ratio calculated based on the current results (OR = 13. 95%CI 3.9 to 44) was highly comparable to the original data (OR = 15, 95%CI 2.1 to 19) using the methods described in the supplementary information of 6.

A more detailed evaluation revealed seven discordant cases between MammaPrint risk assessment and the published data. These cases included two patients that did not develop distant metastases, who were classified as being poor prognosis in the published data 6, but the diagnostic test correctly classified them into the good prognosis group. Furthermore, one patient who did develop metastases was originally classified as good prognosis, whereas in the current results this patient was classified correctly as having a poor prognosis. On the other hand, however, there were two good outcome patients classified as poor prognosis using the diagnostic test, while in the original data these samples were classified correctly, as well as two poor outcome patients classified as good prognosis by the current test who where correctly classified by original analysis as poor prognosis.

To further investigate if the differences seen were due to technical variation of the current test or could be otherwise explained, 49 samples were amplified and hybridized a second time to the 8-pack mini-array (Figure 4). The Pearson correlation for the replicate experiments was 0.995, indicating a very high degree of reproducibility of classification for individual tumor samples using the customized 8-pack array. Also an ANOVA analysis performed on the 70-gene expression values obtained in the duplicate experiments showed no significant differences, independent of variation between individual samples and profile genes (p = 0.960).

To ensure that the outcome of the test does not change over time, two samples were amplified and labeled repeatedly over a period of 12 months (Figure 5). One sample (HRC) was classified as poor prognosis with an average cosine correlation to the good prognosis template of -0.44. The other sample (LRC) was classified as good prognosis (average correlation to the good prognosis template of 0.61). Both samples were stable over time as shown and the diagnostic test result was observed to have a very low standard deviation (LRC and HRC stdev 0.028, i.e. technical variation), indicating the robustness of the diagnostic test.

A sample close to the classification threshold was analyzed 40 times in a period of 4 months. The average correlation to the good prognosis template was 0.430 with a standard deviation of 0.027. The sample was misclassified 6 times (15 %), which is in agreement with the expected chance of misclassification (14%) based on the area of the Gaussian that falls on the other side of the decision boundary (s = 0.028 (the overall standard deviation for the test) and m = 0.430).

To more accurately estimate the risk of metastases associated with the 70-gene prognostic profile, a validation study 13 was performed using a cohort of 295 young breast cancer patients. For the current study we selected the 151 patients from this cohort without lymph node involvement at diagnosis of which 145 RNAs were available.

We calculated the probability of a patient remaining free of distant metastases and overall survival according to the prognosis profile and compared this to the published data 13.

Once more the data generated using the customized array is highly similar (Pearson correlation of 0.88, p < 0.0001) to the original data.

As was seen before, Kaplan-Meier curves showed a significant difference in the probability that patients would remain free of distant metastases when classified in the good or poor prognosis profile group (Figure 6A, LogRank p < 0.001). The difference in prediction of probability of overall survival (Figure 6B) between the groups with good or poor prognosis profiles was also highly significant (p < 0.001). The estimated hazard ratio for distant metastases as a first event in the group with a poor prognosis signature versus the group with a good-prognosis signature, over the entire follow up period, was 5.6 (95% CI 2.4 to 7.3, P < 0.0001). This confirms the published data 13 (HR = 5.5, 95%CI 2.5 to12.2, P < 0.001).

When the probability of a patient remaining free of distant metastases was compared between the current result (Figure 6 blue line) and the original analysis (Figure 6 dashed green line) in the good prognosis profile groups, no significant difference was found (logRank p = 0.890). Similarly for those patients in the poor prognosis profile groups (logRank p = 0.794) (Figure 6 red and dashed magenta lines). Equally, there is no significant difference in overall survival of patients grouped by either the current result or the original published result for either prognosis profile group (two results of good prognosis profile group: logRank p = 0.747, two results of poor prognosis profile group logRank p = 0.760, respectively). All results taken together indicate that there is not only a strong correlation between good prognosis and the absence of distant metastasis or death 13, but the findings generated using the more complex microarray platform were nearly perfectly reproduced using the customized mini-arrays, and demonstrate the robustness of the MammaPrint 8-pack mini-array test.

One hundred and sixty-two patients with lymph node negative (pN0) breast cancer, age of diagnosis before 55 years, not having received adjuvant therapy and that were part of our previous studies 613 were included. All 78 patients that were part of the study in which the 70-gene prognosis profile was established 6, were used to re-establish the 70-gene expression profile. Of the 151 patients used in the clinical validation 13, 145 patients were used, including 61 from the first study. Patients that remained free of disease after initial diagnosis for a period of at least 5 years were assigned to the good prognosis group, i.e. 'good outcome group'; patients that had developed metastasis within 5 years were assigned to the poor prognosis group.

Aliquots of total RNA of 149 frozen tumor samples was available for this study, for 13 samples (8 out of 78 and 5 of the 145 tumor series, see above) new RNA was isolated from available frozen tumor tissue as described previously 61322. Two-hundred nanogram total RNA was amplified using the Low RNA Input Fluorescent Labeling Kit (Agilent Technologies). Cyanine 3-CTP or Cyanine 5-CTP (Perkin Elmer) was directly incorporated into the cRNA during in vitro transcription. A total of 200 ng of Cyanine-labeled RNA was co-hybridized with a standard reference to custom 8-pack mini-microarrays (MammaPrint, Agendia) at 60°C for 17 hrs and subsequently washed according to the Agilent standard hybridization protocol (Agilent Oligo Microarray Kit, Agilent Technologies). The reference sample consisted of pooled and amplified RNA of 105 primary breast tumors selected from patients of the clinical validation series 13 in such a way that it had a similar proportional distribution between good and poor profile patients. Sufficient reference material was generated for over 30,000 hybridizations. For each tumor two hybridizations were performed by using a reversal fluorescent dye.

The customized mini-array contained 1,900 60-mer oligonucleotide probes that comprise the 232 prognosis related genes 6 identical to the probes on the original array, including the genes of the 70-gene prognosis classifier, spotted in triplicate. Each array additionally includes 289 probes for hybridization and printing quality control as well as 915 normalization genes. Eight identical MammaPrint arrays are present on a single 1" × 3" slide, allowing for eight individual hybridizations to be performed simultaneously. After hybridization the slides were washed and subsequently scanned with a dual laser scanner (Agilent Technologies). Microarray raw data are available at the European Bioinformatics Institute (EBI) Arrayexpress database;23 accession number E-TABM-115.

Fluorescence intensities on scanned images were quantified, values corrected for background non-specific hybridization, and normalized using Feature Extraction software version 7.5.1 (Agilent Technologies). Data was further analyzed using custom algorithms in Matlab version 7.1 (The Mathworks). To obtain an overall expression value for each of the signature genes on the array, an error-weighted mean value was calculated for the three identical probes belonging to the same gene as log₁₀ratios. To establish appropriate relative weights, the Rosetta error model was used, which corrects for the uncertainties in individual probe measurements 192425. Probes were excluded from further calculations if their background corrected intensities were below zero and/or if spots were flagged as non-uniformity outliers as determined by the image analysis software.

Outcome prediction for the 78 tumor samples used in Figure 2 and 3 was performed as described by Van 't Veer et al 6. In brief, the 'good prognosis template' was (re-)constructed using the average expression for each of the 70 genes in tumors from the 44 'good outcome' patients as determined on the customized mini-array. Subsequently, the expression of the 70 profile genes for each patient was correlated in a leave-one-out cross validation procedure to the 'good prognosis template'. A patient with a cosine correlation to the good prognosis template higher than 0.4 (the previously determined threshold 6) was assigned to the good-profile group. Patients with a correlation lower than this threshold were assigned to the poor-profile group.

Outcome prediction for the 145 tumor samples used in Figure 6 was performed as described by Van de Vijver et al. (13). For each of the 84 tumors from patients that were not included in the original Nature study 6, a correlation coefficient of the 70-gene expression with the template was calculated as described above. For the 61 patients who were included in the original study 6, correlation coefficients were calculated according to the cross-validated classification method using all 231 genes. This approach was originally employed to minimize to some extent the overestimation of the value of the prognosis profile, i.e., no optimization of the number of reporter genes was carried out, as described in the Nature supplementary information by Van 't Veer et al 6. The only deviation is that 231 instead of a varying number of prognosis correlated genes (range 238 ± 23) were used in the cross-validation procedure since only these 231 genes are present on the mini array. We did show before, however, that the vast majority of the 231 genes were commonly shared by the 78 classifiers generated in the cross-validation procedure 6. For this subgroup, a patient with a tumor with correlation of the 70 genes higher than the previously determined threshold of 0.55 was assigned to the good prognosis profile, and a correlation of less than 0.55 was assigned to the poor prognosis profile group 13.

Odds ratios were calculated based on a two by two contingency table. P-values associated with odds ratios were calculated by Fisher's exact test. Survival periods of patients were analyzed from the calendar date of surgery to the time of the first event or the date on which data were censored, according to the method of Kaplan Meier. The curves were compared using the log rank test.