SELECTING THE GOOD LEVEL OF DETAILS IN UNDECIMATED WAVELET TRANSFORM IMPROVES THE CLASSIFICATION OF SAMPLES FROM METABOLOMIC DATA
Résumé
Wavelet transform is now commonly used to deal with spectrum-shaped metabolomics data. When performing wavelet transform of the initial signal, finest details are often removed before reconstructing the 'true' underlying signal. In this paper, we demonstrate that, on the contrary, finest details may contain relevant information enabling to discriminate sample groups based on their metabolomic profile. We also describe a strategy to determine the best level for classification purpose thanks to a sparse version of PLS-DA. When NMR spectra have the same general shape, removing the coarsest coefficients enables to get rid of the common part of the information and thus to focus on what is truly different between the samples. This strategy is all the more efficient as the level of details is selected in the same step as the discriminant analysis. We obtained NMR spectra from hepatic extracts of wild-type and PPARa-deficient mice exposed to two doses of di-(2-ethylhexyl) phthalate (DEHP), a commonly used plasticizer. The results obtained allowed to easily identify dose- and genotype-dependent changes in hepatic metabolic profiles.