Discovery of survival gene markers for cancer prognosis using genome-wide expression profiles

Abstract

The identification of gene signatures that discriminate between several cancer types using the gene expression profiles derived from genome-wide analyses has been addressed by many authors. Although, gene marker discovery linked to patient prognosis and survival has become increasingly important, it has received less attention and remains a challenging task.In this work, we present a method to identify prognostic gene markers that allow us to partition the sample into groups that maximize the separability between their Kaplan-Meier curves. The separability is evaluated through a trimmed log-rank test. The p-value provided by this test allow us to rank the genes according to their prognostic power. Next, we have developed a methodology to study the association between the best prognostic markers and several clinical variables considered relevant for patient outcome. Finally, the feature selection method has been extended to identify groups of two genes (binary-markers) that relate to the patient prognosis. In particular, the algorithm looks for pair of genes such that both of them change their states in patients of poor outcome.The method proposed has been applied to discover genes associated to response and survival following neoadjuvant taxane-anthracycline chemotherapy for HER2-negative invasive breast cancer. A validation independent cohort of 198 breast cancer patients has been considered in order to evaluate rigorously the algorithms performance. The method is also validated with other cancer series. The experimental results suggest that the feature selection method proposed is able to recover relevant genes in breast cancer prognosis that are missed by other methods.