Stability-based biomarker selection

Biomarker identification, i.e., finding those variables that indicate true differences between two or more populations, is an ever more important topic in the omics sciences. In most cases, the number of variables far exceeds the number of samples, making biomarker identification extremely difficult. We present a strategy based on the stability of putative biomarkers under perturbation of the data, and show that in several cases important gains can be achieved. The strategy is very general and can be applied with all common biomarker identification methods; it also has the advantage that it does not rely on error estimates from crossvalidation, that in this setting tend to be highly variable.