Data Processing Optimization in Untargeted Metabolomics of Urine Using Voigt Lineshape Model Non-Linear Regression Analysis

Abstract

Nuclear magnetic resonance (NMR) spectroscopy is well-established to address questions in large-scale untargeted metabolomics. Although several approaches in data processing and analysis are available, significant issues remain. NMR spectroscopy of urine generates information-rich but complex spectra in which signals often overlap. Furthermore, slight changes in pH and salt concentrations cause peak shifting, which introduces, in combination with baseline irregularities, un-informative noise in statistical analysis. Within this work, a straight-forward data processing tool addresses these problems by applying a non-linear curve fitting model based on Voigt function line shape and integration of the underlying peak areas. This method allows a rapid untargeted analysis of urine metabolomics datasets without relying on time-consuming 2D-spectra based deconvolution or information from spectral libraries. The approach is validated with spiking experiments and tested on a human urine ¹H dataset compared to conventionally used methods and aims to facilitate metabolomics data analysis.

Keywords: NMR; data processing; metabolomics; voigt-fitting.

Figure 1

Step-by-step schematic workflow description for non-linear peak fitting based on Voigt line shape model.

Figure 2

Typical fit results for an exemplary urine spectrum in three regions where signals overlap and/or small peaks are present; initial spectrum is shown as black line, fitted peaks are depicted in colored lines.

Figure 3

Boxplots of standard errors of relative quantification for all three spiked metabolites and methods, individual relative standard errors (RSE) are given as well as the mean RSE ($\overline{\mathrm{RSE}})$ for each method.

Figure 4

Scores plots of PC1 and PC2 using full spectra (A), binned data (B) and Voigt fitted data (C) including 95%-confidence ellipses for each group (Type 2 diabetes mellitus (T2DM) and control); loadings plot for PC1 for all three methods (black) with reference spectrum (blue) (D–F).

Figure 5

Scores plots of predictive and orthogonal variation using full spectra (A), binned data (B) and Voigt fitted data (C) including 95%-confidence ellipses for each group (Type 2 diabetes mellitus (T2DM) and control); loadings plot for first predictive component for all three methods (black) with reference spectrum (blue) (D–F).