Developing gene signatures that are stable, are effective at distinguishing prognostic groups and provide important biological information from whole genome microarray data remains a significant challenge. We propose a method which has similarities to a technique proposed by Bair and colleagues [21
], in combination with an estimation of signature stability [9
] and to our knowledge, the largest dataset of ER+ patients homogenously treated in an attempt to address these issues. Whilst Bair et al. [21
] used the clinical data to define a subset of survival-related genes prior to clustering, we performed an initial unsupervised clustering procedure to form the clusters which could act as biological networks, which were then used as single variables to build the classifier. We hypothesized that this would limit the effect the training set has over the final selection of genes for inclusion in the classifier [4
] and allow a larger gene list for biological hypothesis generating. The inclusion of an assessment of "stability" facilitates determination of the most robust variables and hence presumably important biological information.
With this method, we were able to develop and validate a gene classifier that could predict which patients with ER+ BC were at high risk of relapse despite tamoxifen treatment. Importantly, we were able to validate the classifier on independent samples utilizing raw data from different microarray platforms using a meta-analytical approach. Demonstration of prognostic ability is important if we are to assemble gene lists from microarray data for biological hypotheses generation and potential laboratory experimental validation, which was one of the most important aims of this study. Validation of gene classifiers with independent samples from which they were developed from is a major challenge for microarray studies, especially those with clinical implications, and combining multiple datasets can be difficult due to different patient populations, sample preparation and microarray platforms. Our study uses one of the largest training and validation sets reported in the literature on tamoxifen (only) treated patients.
Whilst, in the future we may have a microarray-based diagnostic test incorporating all 181 genes in the 13 clusters, at present the routine use of this technology is not logistically feasible. However, the advantage of our approach is that as each cluster consists of a group of genes that are highly correlated and hence effectively act as one covariate. Thus, a diagnostic test of just 13 genes (one per cluster) could be developed for clinical use if desired, even though for biological study the researcher would be more interested in all the genes per cluster. To demonstrate this, we took a series of 13 individual probe sets (one per cluster) and correlated their performance with the full classifier on the training set of 255 patients. The median correlation was 0.94 (range: 0.88–0.97). The top 26 ranked 13-gene classifiers (with a correlation ranging from 0.95–0.97) and their corresponding probe sets are listed in [Additional file 7
]. These "simple" tests will require further independent assessment but could be validated using immunohistochemistry or quantitative RT-PCR and are attractive option for potential clinical implementation.
Due to the pressing clinical need, several other investigators have also developed gene predictors that can predict outcome in ER+ BC treated with adjuvant tamoxifen monotherapy [11
]. These studies have used a variety of bioinformatics approaches to develop these gene signatures. These range from a candidate gene approach [24
], selection of genes using a biological approach [23
] and similar to our study, a discovery-based approach using supervised analyses correlated with clinical outcome [13
]. Likewise, different patient populations were used in the development process. Ours is the only study to use a large, consecutive series of patients as a training set as opposed to samples obtained from a clinical trial [24
]; or a case control population [11
]. Only one of these reported gene classifiers has undergone noteworthy clinical validation [24
], however unfortunately these genes provide no new potential therapeutic targets or insights into the underlying biology. Of note, we have previously published that proliferation-related genes are the common biological thread linking many of these currently published classifiers [5
]. Our current classifier also has a significant amount of cell cycle genes, and is highly correlated with the GGI, but one of the aims of this study was to identify other potential biological mechanisms upstream of proliferation. All the clusters in the final classifier were the most common chosen during the cross-validation process suggesting the presence of other strong biological signals. Further experimental validation in in-vitro and in-vivo models will be required to test these hypotheses and their relevance to the clinical question. Interestingly, the cluster 375 was significantly predictive in the dataset of metastatic breast cancer patients treated with tamoxifen as first line treatment for relapsed disease. However, we were not able to validate the full gene classifier. The best approach on distinguishing prognosis versus therapy prediction using gene expression profiling remains unclear. It is possible that developing a predictor of true response to therapy may only be possible using samples from a randomized trial in the metastatic setting where response can be clearly defined and transcriptional profiles can be compared with an untreated control group.