PNNL

Abstract

Metabonomics involves the quantitation of the dynamic multivariate metabolic response of an organism to a pathological event or genetic modification (Nicholson, Lindon and Holmes, 1999). The analysis of these data involves the use of appropriate multivariate statistical methods. Exploratory Data Analysis (EDA) linear projection methods, primarily Principal Component Analysis (PCA), have been documented as a valuable pattern recognition technique for 1H NMR spectral data (Brindle et al., 2002, Potts et al., 2001, Robertson et al., 2000, Robosky et al., 2002). Prior to PCA the raw data is typically processed through four steps; (1) baseline correction, (2) endogenous peak removal, (3) integration over spectral regions to reduce the number of variables, and (4) normalization. The effect of the size of spectral integration regions and normalization has not been well studied. We assess the variability structure and classification accuracy on two distinctly different datasets via PCA and a leave-one-out cross-validation approach under two normalization approaches and an array of spectral integration regions. This study indicates that independent of the normalization method the classification accuracy achieved from metabonomic studies is not highly sensitive to the size of the spectral integration region. Additionally, both datasets scaled to mean zero and unity variance (auto-scaled) has higher variability within classification accuracy over spectral integration window widths than data scaled to the total intensity of the spectrum.