In this article I would like to draw attention to reproducibility of results in biomarker discovery, focusing on the statistical perspective of the problem. Admittedly, reproducibility is a desired property of real markers. However, the important relationship between reproducibility requirement and the stability of statistical feature selection methods is not commonly known yet. So, let me try to clear things up a little...

A need for reproducibility in biomarker discovery

Even though the clinical potential of proteomics and metabolomics in biomarker discovery seems to be high, a non-reproducibility of detected (putative) biomarkers remains the main obstacle. Biomarkers identified by different research groups or even results based on different experiments conducted in the same lab often differ markedly. Hence, there is an emerging need for efficient statistical methods that will address issues of reproducibility and increase our confidence in discovered markers.

Marker reproducibility and feature selection stability

From the statistical perspective, the discovery of biomarkers from high-throughput 'omics' data means searching for the most discriminating features (e.g. features discriminating healthy from disease samples). Such task is usually referred to as feature selection (a good review of feature selection techniques in bioinformatics is given e.g. in [1]).