Statistical correlation based methods for enhanced interpretation of and information recovery from NMR metabolic data sets

Abstract

Owing to its ability to capture a systemic and temporal metabolic description of an organism’s response to a treatment, metabonomics is a well-established and valuable approach in elucidating the effects and mechanisms of a given perturbation. However, to optimise information recovery from the complex datasets generated, chemometric methods are essential. The work presented in this thesis focuses on the development of novel methods, and the use of existing methods in new applications to ease data interpretation and enhance information recovery from 1H Nuclear Magnetic Resonance (NMR) metabonomic datasets using correlation based methods. Although the methods here are largely applied to toxicological data, they could be equally valuable in the analysis of any metabonomic dataset, and indeed potentially to other ‘omics’ data presenting similar analytical challenges. The first two methodological approaches relate to novel extensions of Statistical Total Correlation Spectroscopy (STOCSY), a valuable tool in elucidation of both inter- and intra-metabolite spectral intensity correlations in NMR metabonomic datasets. In the first, STOCSY is utilised in STOCSY-editing, a method for the selective identification and downscaling of the peaks from unwanted metabolites such as those arising from xenobiotics. Structurally correlated peaks from drug metabolites are first identified using STOCSY, and the returned correlation information utilised to scale the spectra across these regions, producing a modified set of spectra in which drug metabolite contributions are reduced, endogenous peaks reconstructed and thus, analysis by pattern recognition methods without drug metabolite interferences facilitated. In the second, the STOCSY approach is extended in Iterative-STOCSY, where metabolic associations are followed over several rounds of STOCSY through calculation of correlation coefficients initially from a driver spectral peak of interest, and subsequently from all peaks identified as correlating above a set threshold to peaks picked in the previous round. The condensation of putatively structurally related peaks into single nodes, and representation of the otherwise complex network in a fully interactive plot of node-to-node connections and corresponding spectral data, allows the ready exploration of both inter- and intrametabolite relationships and a more directed approach to the identification of biomarkers of the studied perturbation. Finally various clustering methods are investigated with the aim of providing improved structural (intra-metabolite) versus non-structural (inter-metabolite) assignment. Thus, this thesis presents a framework for the enhanced identification, recovery and characterisation of inter- and intrametabolite relationships and how these are affected by metabonomic perturbation.