Feature selection is critical in omics data analysis to extract restricted and meaningful molecular signatures from complex and high-dimension data, and to build robust classifiers. This package implements a new method to assess the relevance of the variables for the prediction performances of the classifier. The approach can be run in parallel with the PLS-DA, Random Forest, and SVM binary classifiers. The signatures and the corresponding 'restricted' models are returned, enabling future predictions on new datasets. A Galaxy implementation of the package is available within the Workflow4metabolomics.org online infrastructure for computational metabolomics.