Robust gene expression signature is not merely a significant P value

Since the seminal publication in 1999, the role of gene expression signatures in oncology has grown, particularly for predicting outcome or response to chemotherapy. Nevertheless, legitimate excitement about the attractiveness of molecular technologies and the promise of discovery-based research should not overlook adherence to the rules of evidence, otherwise it may result in claims that are not meaningful and lead to disappointment. The very first of these rules has to be the significance of the findings, both at the statistical and biological levels. Few studies based on the outcomeassociation argument report negative controls to check whether their signature of interest is actually more strongly related to outcomes than signatures with no oncological rationale. Faced with this fact, Venet et al. recently reported a very interesting study aiming to test the null hypothesis assuming a background of no associations with outcomes. They compared 47 published breast cancer outcome signatures to signatures of identical size constructed from random genes and showed that 60% of the published signatures were not significantly better outcome predictors than random signatures. Here we report exactly the same approach regarding the CINSARC signature recently published and validated