Combining Spectral Ordering with Peak Fitting for One-dimensional NMR Quantitative Metabolomics

One-dimensional (1)H NMR spectra are widely used for metabolic profiling. Such data sets often contain hundreds or thousands of spectra, which typically have variation in their sample chemistry, which leads to chemical shift variation "positional noise". This is a severe problem for metabolite quantification and data analysis, as peak integrals do not necessarily correspond across all spectra in a set. Various alignment algorithms have been developed to address this problem, but different studies have taken different approaches to evaluating the performance of NMR alignment routines and can be subjective or rely on an arbitrary cutoff. Furthermore, most alignment approaches completely fail to deal with peaks that overlap. We adopt the simple and robust method of ordering spectra with respect to an internally varying peak and use this to compare different alignment algorithms. Furthermore, we use the information from this procedure to help improve a Bayesian approach to automated peak deconvolution by restricting the prior probability distribution of the peak position in a model-free manner and compare the performance to manual peak deconvolution and to binning. This combination of spectral ordering and compound deconvolution improved the quality of the data for quantitative metabolomics.