We collected publicly available datasets from journal articles and repositories such as Gene Expression Omnibus (GEO) and ArrayExpress, selecting those with a medium to large sample size (Table 1). Since publications sometimes used the same patients, datasets with unique patients were introduced (identified by the 'dataset symbols' in Table 1) by merging some original datasets or removing redundant patients. The collection includes datasets produced on whole-genome microarrays, small diagnostic arrays, and reverse transcription-polymerase chain reaction panels, totaling 2,833 expression profiles. Hybridization probes were mapped to Entrez GeneID 26 through sequence alignment against RefSeq mRNA in the (NM) subset, similar to the approach of Shi and colleagues 27, using RefSeq version 21 (2007.01.21) and Entrez database version 2007.01.21. When multiple probes were mapped to the same GeneID, the one with the highest variance in a particular dataset was selected to represent the GeneID. The numbers of distinct GeneIDs obtained for each dataset are shown in Table 1. The normalized, log-transformed expression measures as published by the original studies were used. Meta-analyses were performed on the union of all 17,198 genes. Summary statistics of absent genes were considered as missing values. Summaries of the availability and compositions of important clinical variables for each dataset are shown in Figure 1 of Additional data file 2.

The expression levels of the prototype genes on the log₂ scale were used as explanatory variables in multiple regression with the Gaussian error model, using the following equation (gene symbols stand for their log expression, and coefficients are omitted for clarity):

Y_i = ESR1 + ERBB2 + AURKA,

where the response variable Y_i is the expression of gene i. This model is fitted separately for each gene i in the array. The association between gene i and prototype j (in the presence of or conditional on all other prototypes) is tested using the t statistic for each coefficient. Because the t statistics for different datasets have different degrees of freedom, we put them all on the same scale by transforming them to the corresponding cumulative probabilities and then to z scores using the inverse standard normal cumulative distribution function. The z scores were combined meta-analytically across datasets using the 'inverse normal method'. The linear model above was fitted separately to each gene in each dataset, and the z scores were combined meta-analytically over multiple studies using the inverse normal method 28. To select genes that are most strongly associated with the prototypes, we use a stringent criterion of |z| ≥ 16, which is well above |z| ≈ 5 that corresponds to a Bonferroni-corrected P value of 0.05.

For a specific dataset, the module score is computed for each sample as

m o d u l e s c o r e = ∑ i w i x i / ∑ i | w i | , MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuqqRPxAKvMB6bYrY9gDLn3AGiuraeXatLxBI9gBaebbnrfifHhDYfgasaacPi6xNi=xI8qiVKIOFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaqefCuAVzxyU5wAGi0BVTgaiqaacqWFTbqBcqWFVbWBcqWFKbazcqWF1bqDcqWFSbaBcqWFLbqzcqWFGaaicqWFZbWCcqWFJbWycqWFVbWBcqWFYbGCcqWFLbqzcqGH9aqpdaaeqbqaaGabciab+Dha3naaBaaaleaacqWGPbqAaeqaaOGae4hEaG3aaSbaaSqaaiabdMgaPbqabaaabaGae4xAaKgabeqdcqGHris5aOGaei4la8YaaabuaeaacqGG8baFcqGF3bWDdaWgaaWcbaGae4xAaKgabeaakiabcYha8bWcbaGae4xAaKgabeqdcqGHris5aOGaeiilaWcaaa@58B2@

where x_i is the expression of a gene in the module that is present in the dataset's platform. w_i is either +1 or -1, depending on the sign of the z score of the association with the prototypes.

To cluster the tumors based on the ESR1 and ERBB2 module scores, Gaussian mixture models 29 with equal and diagonal variance for all clusters were fitted. For testing multimodality, we used the likelihood ratio test statistics between the fitted model for the tested number of components, k, versus the alternative model with k - 1 components. The statistical significance of the number of components was assessed by parametric bootstrapping. Each tumor was automatically classified as estrogen receptor-negative (ER^-)/HER2⁺, HER2⁺, or ER⁺ using the maximum posterior probability of membership in the clusters.

Survival curves and 5-year survival rates in forest plots were based on Kaplan-Meier estimates, with the Greenwood method used for computing the 95% confidence intervals 30. Hazard ratios between two groups were calculated using Cox regression. Stratified Cox regression was used to compute total hazard ratios in forest plots and multivariate analysis, using the dataset as the stratum indicator, thus allowing for different baseline hazard functions between cohorts. Cox regression was also used to compute gene-by-gene associations with survival, treating the log expression measures as continuous explanatory variables. The gene-wise z scores were combined across datasets using the inverse normal meta-analytical methods. Distant relapse-free survival (DRFS) was considered as an event for our survival analysis, which includes distant recurrence, death from breast cancer, death from a cause other than breast cancer, and death from an unknown cause.

To perform this meta-analysis including several heterogeneous datasets and different microarray platforms, we used the concept of coexpression modules. To identify these modules, we applied a supervised approach whereby three 'prototype' genes representing three key biological processes in breast cancer (namely, proliferation, ER, and HER2 amplification signaling) were selected. The genes chosen as their prototypes were, respectively, ESR1, ERBB2, and AURKA (aurora-related kinase 1, also known as STK6 or STK15).

Using the meta-analysis scheme described above, we were able to identify genes whose expression was significantly associated with each chosen prototype (Additional data file 3). The coexpression patterns of the genes are shown by heatmaps in Figure 2 of Additional data file 2. Each module contains highly correlated or anticorrelated genes, as shown by the vertical color patterns. The annotation of the modules shows that they correspond well to the expected biological processes, as many ER-related, HER2-related, and proliferation genes were included in the ER and HER2 signaling and proliferation modules, respectively. For our further analysis, the correlated gene expression measures in a module (which provide redundant information) are averaged into a single number called a 'module score'.

To automatically assign these large numbers of tumors into the subtypes according to the given module, we applied the Gaussian mixture models 29 to the module scores of the three processes. Only three natural clusters, based on multimodality tests, can be identified. The ER and HER2 module scores were bimodally distributed, but the proliferation module was not. Furthermore, the combination of the ER and HER2 module scores does not produce four clusters that would have been observed if the scores were independent (Figure 1a). Instead, ERBB2⁺ tumors showed an intermediate level of ER module values, and we therefore did not consider the distinction of ERBB2⁺ into ER⁺ or ER^- to be supported by continuous value gene expression levels. We will refer to three groups as ER^-/HER2^-, HER2⁺, and ER⁺/HER2^- tumors, which correspond roughly to the intrinsic subtypes of basal-like, her2, and combined luminal A/B subtypes, respectively, as defined by the Stanford group 15.

Concerning proliferation, Figure 1b shows that, while ER^-/HER2^- and HER2⁺ tumors have mostly high proliferation scores, ER⁺/HER2^- tumors display a wide range of values, encompassing the low values of normal breast tissue (see dataset UNC) and the high values typical for ER^-/HER2^- and HER2⁺ tumors. For our further analysis, we denote the ER⁺/HER2^- low- and high-proliferation tumors as ER⁺/HER2^-/L and ER⁺/HER2^-/H, corresponding to the luminal A and B subdivisions of the intrinsic subtypes, respectively. Interestingly, we did not see natural clustering (bimodality) in the distribution of proliferation scores as was the case with the ER and ERBB2 modules.

The relationship between module scores and some gene mutations could also be examined. Almost all BRCA1-mutated tumors are confined to ER^- tumors (Figure 1b), confirming the hypothesis that ER^- ('basal-like') tumors are phenocopies of BRCA1-mutated tumors 14. This is also supported by the strong overexpression of LMO4, a suppressor of BRCA1 function 31, in ER^- tumors. p53 mutations may appear in the three subtypes, but mostly confined to the highly proliferative tumors. It is not clear whether their association with ER^-/HER2^- and HER2⁺ tumors is related to the pathways of these receptors or is merely an indirect effect of the mutations' association with proliferation.

The attractiveness of gene expression prognostic signatures for clinical applications comes from their ability to identify a group of patients with a good survival rate that is acceptable to spare patients from aggressive chemotherapy. Here, we investigated whether classifications based on the easily interpretable module scores could achieve such clinical relevance.

Figure 2 shows a Kaplan-Meier analysis for the DRFS of systemically untreated (Figure 2a) patients and those treated (Figure 2b) with adjuvant chemotherapy and/or endocrine therapy with available clinical information, according to four main subtypes based on the module scores. The ER⁺/HER2^-/L subtype showed a much better DRFS than the three others in both untreated and treated populations, with 90% of patients alive at 5 years of follow-up. Because there is no statistical difference in survival between the ER^-, HER2⁺, and ER⁺/HER2^-/H subtypes and because the risk of recurrence for patients in these groups is clinically still too high, we pooled them into the 'poor' prognosis group, in contrast to the 'good' ER⁺/HER2^-/L subtype, for further survival analysis. The consistency of the prognostic value across datasets is demonstrated by the forest plots in Figures 2c and 2d, where the results of the analysis of individual datasets are concisely summarized by the 5-year survival estimates and hazard ratios between the 'good' and 'poor' groups. Interestingly, the 'good' prognosis group showed a better DRFS than the 'poor' prognosis group in both untreated and systemically treated populations.

The interactions between the module-based risk groups and conventional clinicopathological prognostic variables are tested in multivariable Cox regression analysis for DRFS in both untreated (Figure 2e) and treated (Figure 2f) populations. The module-based classification added a strong prognostic effect over all other clinical factors. Confirming previous studies 1832, the effect of histological grade is much reduced and can be explained by the refinement of intermediate grade into two groups with very different survival rates. Interestingly, lymph node status and tumor size remain as independent prognostic factors.

Although Fan and colleagues 25 noted the similarity of the performance and patient classifications of the intrinsic subtypes and four prognostic signatures on the same dataset, they did not provide a biological rationale for this finding. In our study, we performed more detailed and extensive analysis to better understand how disparate gene lists may give rise to potentially equivalent prognostic signatures.

Using our meta-analytical approach, we first sought to identify individual genes that were associated with survival by calculating the meta-analytical z scores of gene-by-gene Cox regression. To gain further insight into the biological significance of these prognostic genes, we investigated their correlation with the coexpression module prototypes. We were able to identify 524 genes that were significantly associated with survival, even under a stringent Bonferroni multiple testing correction (data not shown). Of the 524 genes, 71% were strongly coexpressed with proliferation, 26% with ER, and 2.2% with ERBB2 prototypes, highlighting the importance of proliferation-related genes for prognostication in breast cancer.

A similar analysis was performed with respect to several published prognostic signatures (Table 2). Indeed, many of the genes included in these signatures were confirmed to be individually prognostic in the whole dataset collection (Figure 3 of Additional data file 2). Interestingly, many of these individually prognostic genes were also highly correlated with the proliferation module prototype and not with the other two modules, suggesting that proliferation may be the common driving force of several prognostic signatures.

To further demonstrate our hypothesis, we divided each signature into two 'partial signatures': one with only proliferation genes and the other with the complementary nonproliferation genes (Figure 3; see Figure 4 of Additional data file 2 for detailed analysis). Interestingly, when only proliferation genes were used, the overall performance was not degraded; in fact, it even improved for some signatures (p53-32) in both untreated (Figure 3a) and treated (Figure 3c) populations. In contrast, the nonproliferation partial signatures typically showed degraded performance (Figures 3b and 3d). These results show that proposed signatures may contain genes that are unnecessary or even detrimental to their performance. These results thus extend the findings of Fan and colleagues 25 to a much larger sample size and for several additional signatures, revealing for the first time the importance of proliferation genes as a common driving force behind the performance of all of the prognostic signatures studied in this investigation.

Finally, the relationship between prognostic signatures and the molecular classification based on the coexpression modules was investigated by looking at the risk classifications on the plots of proliferation scores versus the molecular subtypes shown in Figure 4 (see Figure 4 of Additional data file 2 for analysis on all datasets). Most signatures identified the low-proliferation subset of ER⁺/HER2^- tumors as low-risk, whereas almost all high-proliferation ER⁺, ER^-/HER2^-, and HER2⁺ tumors were classified as high-risk. These results suggest that these prognostic signatures function mostly by identifying tumors that have high expression of proliferation genes, regardless of the subtyping based on ER or HER2. They still correctly classify ER^-/HER2 and HER2⁺ as high-risk by virtue of elevated expression of proliferation genes.