Variable selection for a mixed population applied in proteomics

The paper presents a variable selection method for biomarker discovery in proteomics. More specifically, it finds the most adequate variables among a given set in order to discriminate between two groups (healthy and pathological). This approach is developped within a Bayesian framework and relies on an optimal strategy that results in the choice of the most a posteriori probable model. The calculation of the posterior probabilities requiresmarginalization of unknown parameters. It is the main difficulty and a contribution of the paper is to provide a closed-form expression. The originality of the work is twofold: (1) we relax the standard hypothesis of linear regression models and (2) we present a multivariate test which directly accommodates possible correlations between the biomarkers. The effectiveness of the method is assessed through a simulated study and shows results in accordance with the theoritical optimality.