Many studies have demonstrated the ability of using gene-expression "signatures" derived from DNA microarray data to define cancer subtypes, predict disease recurrence, and guide treatment decisions. In breast cancer, van't Veer et al. [1] derived a 70-gene profile to predict a patient's risk of developing distant metastases. Perou et al. [2] and Sorlie et al. [3] developed an intrinsic-subtype signature that classifies breast tumors into molecular subtypes showing distinct differences in prognosis. From a cancer biology perspective, Chang et al. [4] studied the links between the wound healing process and cancer progression. Based on the expression pattern of a wound-response signature of 512 genes, they classify a tumor to have either activated or quiescent response and found this to be a significant prognostic predictor of tumor metastasis. These are promising results and a few of these signatures have begun to be assessed in clinical settings. Two questions have often been asked: (1) are these signatures identifying the same set of genes and (2) will they generate similar prediction performance when tested in new data sets?

The answer to the first question has been discouraging. Any pair of these signatures share only a few common genes. Possible reasons have been suggested including the differences in patient cohort characteristics (such as the distribution of age or stage of the disease), lack of comparability and reproducibility of the data generated using different microarray platforms, and varying statistical procedures used to generate the gene list. Nevertheless, Ein-Dor et al. [5] showed that the inconsistency still exists when eliminating all three differences. In particular, the authors repeated the same analysis in a single data set and identified many lists of genes equally predictive of the outcome. Any two of these gene lists share only a small number of genes. In another study by Son et al. [6], the authors reported that any randomly selected subgroup of around 100 differentially expressed genes generates similar hierarchical clustering results in the same data set.

Ein-Dor et al. [7] further suggested perhaps the main source of the problem lies in the small sample size and large number of genes the signatures were derived from. For several published breast cancer data sets, the authors estimated that several thousands of samples would be needed to achieve a typical gene overlap of 50%. On the other hand, the problem is compounded by analyzing and interpreting genes in isolation. A common approach to gene selection involves selecting a handful of top-ranking genes that best differentiate sample classes (such as tumor vs. normal tissue) or are most predictive of clinical outcome. The univariate selection procedure ignores correlation between genes. The biological and statistical validity of such assumption seems tenuous. As a result, gene-set based approaches have emerged in recent years to identify sets of biologically related genes that are deregulated as a group. Examples of gene-set analysis include the Gene Set Enrichment Analysis (GSEA) [8], Significance Analysis of Function and Expression (SAFE) [9], and the globaltest package [10] These methods focus on groups of genes that share common biological functions such as cell cycle regulation; metabolic or signaling pathways defined by Gene Ontology (GO); online databases such as BioCarta, KEGG and signaling data base; or a literature-defined gene set subject to experimental perturbations such as a drug treatment or an oncogene-activation. In addition, Rhodes et al. [11] introduced a Molecular Concepts Map (MCM) providing an expanded analytic framework to explore the network of relationships among biologically related gene sets.

The motivation of this study came from a recent paper by Fan et al. [12], which addresses the second question described above. The authors demonstrated a high degree of prediction concordance of five breast cancer gene-signatures despite minimal gene-wise overlap. In an independent data set of 295 tumors, the authors showed that the intrinsic subtypes [3] of basal-like, HER2+/ER-, and luminal B were consistently classified as poor 70-gene profile [13] prognosis, activated wound response [4] and high recurrence score [14]. It raises the question that perhaps the gene-overlap is not the most relevant measure of robustness and reproducibility of the gene-signatures. There may be common themes shared across these signatures that are not apparent on the individual gene level. As an example, the cell cycle gene Cyclin E1 (CCNE1) was included in the 70-gene profile while Cyclin E2 (CCNE2) in the intrinsic subtype signature. The two signatures apparently share commonality in the activation of the Cyclin family genes. For another example, ERBB2 and EGFR are both receptor tyrosine kinase involved in estrogen pathway. Inclusion of one or the other in two different signatures apparently converges at the pathway level both indicating the activation of the estrogen-signaling pathway.

In this study, we assess the potential functional convergence of these gene-signatures on the basis of activated oncogenic pathways. This involves first annotating each gene-signature to identify significantly enriched functional modules (e.g., cell growth, response to estrogen, myb-regulated pathways, etc.). Definition of the modules can be based on Gene Ontology (GO) terms, online pathway databases such as BioCarta and KEGG, or literature-defined concepts. In the next step, the overlapping functional modules are obtained by intersecting the annotated sets. We investigated six breast cancer signatures (four of which were compared in Fan et al. [12]) that share high prediction concordance. We found eighteen common modules including estrogen-signaling, responses to tamoxifen treatment, and BRCA1 expression. The degree of the functional overlap across the six BR-signatures is highly significant (P = 0.0002) under a bootstrapped null distribution.

Cancer gene-expression signatures derived from microarray experiments are beginning to be tested in clinical trials, while the exact biology that enables these gene-signatures to accurately predict tumor metastasis and patient survival is unclear. Microarray experiments are often limited in power by the small number of samples used to derive a panel of prognostic genes relative to the large number of features on the array. In addition, sets of biologically-related genes are often co-regulated while many feature selection procedures are univariate in nature. As a result, gene-signatures developed by different studies typically share very few common components. A recent study showed high prediction concordance of several breast cancer gene-signatures despite minimal overlap in gene identity. It gave main motivation to investigate common oncogenic themes that may not be apparent at the individual gene level. This study explored this hypothesis by evaluating the functional overlap of the signatures on the basis of annotated gene sets. When the gene signatures are mapped to the deregulated pathway space, two things become clear. First, there is a significant degree of functional overlap in oncogenic and prognostic pathways. Second, many of these common pathways provide plausible explanations of tumor biology through which these signatures predict patient outcome. There are several conclusions to be gleaned from this study. First, this work explains why independent signatures that appear to perform equally well at predicting patient prognosis show minimal overlap in gene membership. This is because such genes are different members of pathways and processes that are relevant to prognosis. Thus, the lack of gene overlap found between the various signatures listed in Table 1 should not be considered problematic. The implication of our study is that most of these signatures will do well in clinical trials given that they seem to be picking up the same pathway signals. We can thus be assured that the gene lists found by different investigators are consistent, even if they do not contain the same genes.

Second, the results have suggested that the interpretability and delineation of how diverse cancer gene expression signatures work are more likely attainable at the pathway level rather than the individual gene level. On the other hand, as many studies have already suggested so, feature selection methods need to be based on biologically related gene sets that are deregulated as a group [8–11]. However, it is not a straightforward task to construct a prognostic signature based on pathways that are composed of overlapping sets of genes. New statistical methods need to be established in this area. This is beyond the scope of the study and is currently under investigation.

Table 1 lists the six BR-signatures that are compared in this study. Fan et al. [10] showed high prediction concordance of signature 1–3 and signature 5. In addition, a 90-gene meta-signature [13] is included. This signature was derived in a meta-analysis framework by integrating four microarray data sets, which included the van't Veer data set and the Sorlie data set. Another signature included here is the subset of 60-gene profile from Wang et al. [18] that was derived in tumors with estrogen receptor (ER) positive status. The recurrence-score signature [14] is also an ER+ disease signature that has been shown in a clinical trial to be able to identify patients with very low risk of recurrence on hormone therapy using tamoxifen alone, and do not require adjuvant chemotherapy.