| Metabolomics is the comprehensive study of the chemical fingerprint of specific cellar metabolic processes. Nuclear magnetic resonance spectroscopy (NMR) is one major detection method in identifying and quantifying metabolites in metabolomics studies. The data analysis in NMR metabolomics spectra is a complex task involving many computational and statistical procedures. The most important features of NMR metabolomics datasets are high dimensionality and peak displacements under experimental variations. This dissertation focuses on integrating, improving and developing data processing techniques to extract metabolites information for metabolic profiling and biomarker discovery specifically via NMR spectroscopy. We developed an improved baseline correction method based on penalized parametric smoothing model and an adaptive binning method based on peak density clustering in spectra preprocessing, and compared two dimension reduction methods as well as three classification methods to find the optimal combined approach for detecting significant variables in the study of a genetic disease, pseudoxanthoma elasticum (PXE). We also developed two fast screening methods in compound identification in 2D COSY and HSQC spectra by statistically modeling the peak displacements under pH variations. The theoretical methodologies, numerical implementations, parameter optimization and results for these methods are presented. |
