Combination of Independent Component Analysis and statistical modelling for the identification of metabonomic biomarkers in 1H-NMR spectroscopy (second version)

In order to maintain life, living organisms product and transform small molecules called metabolites. Metabonomics is a recent scientic platform, studying the development of biological reactions caused by a contact with a physio-pathological stimulus, through these metabolites.The 1H-NMR spectroscopy is widely used to graphically describe a metabolites composition via spectra. Biologists can then conrm or invalidate the development of a biological reaction if specic NMR spectral regions are altered from a given physiological situation to another. However, this process supposes a preliminary identication step which traditionally consists in the study of the two rst components of a Principal Component Analysis (PCA). This paper presents a new methodology in four steps providing knowledge on specic 1H-NMR spectral areas (and by extension on biomarkers) via the identication of biomarkers as such, and via the visualization of the eects caused by some external changes. A rst step implies Independent Component Analysis (ICA) in order to decompose the spectral data into statistically independent components or sources of information. The independent (pure or composite) metabolites contained in biouids are discovered through the sources, and their quantities through mixing weights. The advantages of independent components are described in comparison with usual PCA analysis. Specic questions related to ICA like the choice of the number of components or their ordering will be discussed here. The second step consists in a statistical modelling applied to the ICA outputs. The third step will introduce statistical hypothesis tests on the parameters of the estimated models, with the objective of selecting sources which present biomarkers or signicantly uctuating spectral regions according to our factor of interest. A panel of various statistical models is considered here, that can adapt to dierent possible kinds of data or dierent investigations. Finally, the last step proposes a computation of contrasts which can lead to the visualization of changes on spectra caused by changes of the factor of interest. The whole methodology is illustrated on two experimental datasets