The influence of feature selection methods on accuracy, stability and interpretability of molecular signatures

Anne-Claire Haury (1, 2) , Pierre Gestraud (2) , Jean-Philippe Vert (1, 2)

1 CBIO - Centre de Bioinformatique
2 Cancer et génome: Bioinformatique, biostatistiques et épidémiologie d'un système complexe

Motivation : Biomarker discovery from high-dimensional data is a crucial problem with enormous applications in biology and medicine, such as, e.g. breast cancer prognosis. While it is now common belief that feature selection methods should output both accurate and stable signatures - at least at the functional level -, no study has focused on comparing algorithms in these regards. Methods : Borrowing 4 public datasets, we define systematic procedures to compare a representative panel of 8 feature selection algorithms, among which filters, wrappers and embedded methods in light of predictive performance, stability and functional interpretability of the signatures that they output. We also implemented Ensemble methods and propose to estimate the advantages of using them. Results : We observe that ensemble feature selection techniques have generally no substantial impact on accuracy or stability. Additionally we notice a possible trade-off between stability and accuracy as some methods produce predictive but unstable signatures while others behave in the opposite way. Filter feature selection by a Student's t-test seems to provide a good accuracy/stability trade-off overall.

Mots clés: biomarker discovery; gene expression; microarrays; feature selection; signature; breast cancer

Domaines: Bio-Informatique, Biologie Systémique [q-bio.QM]; Bio-informatique [q-bio.QM]; Machine Learning [stat.ML]; Applications [stat.AP]

Anne-Claire Haury, Pierre Gestraud, Jean-Philippe Vert. The influence of feature selection methods on accuracy, stability and interpretability of molecular signatures. 2010. ⟨hal-00559580v1⟩