Processing and modeling of nuclear magnetic resonance (NMR) metabolic profiles

Modern nuclear magnetic resonance (NMR) spectroscopy generates complex and information-rich metabolic profiles. These require robust, accurate, and often sophisticated statistical techniques to yield the maximum meaningful knowledge. In this chapter, we describe methods typically used to analyze such data. We begin by describing seven goals of metabolic profile analysis, ranging from production of a data table to multi-omic integration for systems biology. Methods for preprocessing and pretreatment are then presented, including issues such as instrument-level spectral processing, data reduction and deconvolution, normalization, scaling, and transformations of the data. We then discuss methods for exploratory modeling and exemplify three techniques: principal components analysis, hierarchical clustering, and self-organizing maps. Moving to predictive modeling, we focus our discussion on partial least squares regression, orthogonal partial least squares regression, and genetic algorithm approaches. A typical set of in vitro metabolic profiles is used where possible to compare and contrast the methods. The importance of validating statistical models is highlighted, and standard techniques for doing so, such as training/test set and cross-validation are described. Finally, we discuss the contributions of statistical techniques such as statistical total correlation spectroscopy, and other correlation-based methods have made to the process of structural characterization for unknown metabolites.