Evaluation of standard and advanced preprocessing methods for the univariate analysis of blood serum 1H-NMR spectra

Abstract

Proton nuclear magnetic resonance ((1)H-NMR)-based metabolomics enables the high-resolution and high-throughput assessment of a broad spectrum of metabolites in biofluids. Despite the straightforward character of the experimental methodology, the analysis of spectral profiles is rather complex, particularly due to the requirement of numerous data preprocessing steps. Here, we evaluate how several of the most common preprocessing procedures affect the subsequent univariate analyses of blood serum spectra, with a particular focus on how the standard methods perform compared to more advanced examples. Carr-Purcell-Meiboom-Gill 1D (1)H spectra were obtained for 240 serum samples from healthy subjects of the Asklepios study. We studied the impact of different preprocessing steps--integral (standard method) and probabilistic quotient normalization; no, equidistant (standard), and adaptive-intelligent binning; mean (standard) and maximum bin intensity data summation--on the resonance intensities of three different types of metabolites: triglycerides, glucose, and creatinine. The effects were evaluated by correlating the differently preprocessed NMR data with the independently measured metabolite concentrations. The analyses revealed that the standard methods performed inferiorly and that a combination of probabilistic quotient normalization after adaptive-intelligent binning and maximum intensity variable definition yielded the best overall results (triglycerides, R = 0.98; glucose, R = 0.76; creatinine, R = 0.70). Therefore, at least in the case of serum metabolomics, these or equivalent methods should be preferred above the standard preprocessing methods, particularly for univariate analyses. Additional optimization of the normalization procedure might further improve the analyses.