Comparison of some chemometric tools for metabonomics biomarker identification

NMR-based metabonomics discovery approaches require statistical methods to extract, from large and complex spectral databases, biomarkers or biologically significant variables that best represent defined biological conditions. This paper explores the respective effectiveness of six multivariate methods: multiple hypotheses testing, supervised extensions of principal (PCA) and independent components analysis (ICA), discriminant partial least squares, linear logistic regression and classification trees. Each method has been adapted in order to provide a biomarker score for each zone of the spectrum. These scores aim at giving to the biologist indications on which metabolites of the analyzed biofluid are potentially affected by a stressor factor of interest (e.g. toxicity of a drug, presence of a given disease or therapeutic effect of a drug). The applications of the six methods to samples of 60 and 200 spectra issued from a semi-artificial database allowed to evaluate their respective properties. In particular, their sensitivities and false discovery rates (FDR) are illustrated through receiver operating characteristics curves (ROC) and the resulting identifications are used to show their specificities and relative advantages.The paper recommends to discard two methods for biomarker identification: the PCA showing a general low efficiency and the CART which is very sensitive to noise. The other 4 methods give promising results, each having its own specificities.