Nuclear magnetic resonance (NMR) spectroscopy is a common method to analyze metabolites in biological fluids such as urine or blood plasma [
1,
2]. The resulting spectra consist of a large number of signals at different frequency positions (
x-axis), at which each peak corresponds ideally to a certain metabolite, while its signal volume reflects its concentration in the fluid. However, due to the large number of metabolites present in a typical biological sample substantial signal overlap is commonly observed. Therefore, a precise discrimination between individual overlapping metabolite signals is often not feasible, hampering both signal assignment and accurate quantification. To overcome this drawback several methods have already been developed. One obvious solution is to spread the overlapping signals over more than one dimension [
3,
4]. However, this comes, in the case of
H-
C HSQC spectra measured at natural abundance, at the price of reduced sensitivity and prolonged measurement time, which may be partly compensated by the application of non-uniform sampling schemes [
5] or by application of so-called ultrafast NMR methods, where the conventional time incrementation has been replaced by spatial encoding [
6]. Consequently, for large studies comprising up to several thousand samples or in cases where NMR measurement time is limiting, 1D
H NMR spectra will be commonly used. Therefore, to obtain accurate quantitative information from 1D spectra containing signal overlap, spectral deconvolution techniques are required. For example, Gaussian or Lorentzian line shapes may be fitted to
P NMR spectra to analyze single compounds such as inositol phosphates [
7]. For the analysis of 1D fluorine spectra of proteins, Hughes et al. developed an approach employing Bayesian information criteria to objectively determine the minimal number of signals required to reproduce the experimental data [
8]. In the case of the simultaneous analysis of multiple compounds, one widely used approach is to fit reference spectra of standards to approximate the measured mixture spectrum. This can be done in a semi- or fully-automated fashion using, for example, the commercial software Chenomx (Chenomx Inc. Edmonton, Canada), or freely available tools such as Bayesil [
9] or Batman [
10]. Common to these methods is their requirement of reference information obtained for pure compounds. Moreover, they do not provide quantitative information for NMR signals of yet unknown identity. A recent approach is the SigMa software [
11], which was developed for the analysis of urinary spectra and which is able to obtain quantitative information from both known and unknown metabolites. To this end, SigMa focuses mainly on the analysis of spectral intervals representing only single compounds, while for spectral regions containing signal overlap a binning approach is used.
Spectral deconvolution including a combination of time-frequency analysis and probabilistic sparse matrix factorization has been used as a preprocessing step to reduce noise in NMR spectra [
12]. For 2D
H-
C HSQC spectra, Chylla et al. implemented a deconvolution algorithm based on the fast maximum likelihood reconstruction (FMLR) [
13] to provide accurate signal integrals without the use of reference spectra. As an advancement of the previously developed Batman package, a Bayesian deconvolution algorithm for the automated analysis of 2D JRES spectra has been introduced [
14]. Additionally, in in vivo NMR analyses spectral deconvolution is an important topic as drastically increased line widths result in a substantial amount of signal overlap. Over the years numerous approaches such as LCModel [
15] for metabolite quantitation in in vivo NMR spectra have been developed. An overview of existing approaches was recently given by Barker et al. [
16]. Additionally, for high-dimensional metabolomic data generated by means of liquid chromatography coupled-mass spectrometry (LC-MS), fully automated signal deconvolution methods using continuous wavelet transforms were developed [
17].
The aim of the present contribution was the development of an easy to use
R-package for the deconvolution of overlapped signals in 1D NMR spectra without the need for reference spectra. To this end, a fully automated determination of the underlying Lorentzian lines, which is the natural line-form of NMR signals, was implemented. The method is based on previous theoretical work by Koh et al. [
18,
19], which we adapted to optimally work with highly complex spectra of human biofluids such as urine and plasma. Our approach is also distantly related to work by Schmidt et al. who used Lorentzian line shapes for the deconvolution of electrophoretic NMR data to study non-aqueous electrolytes [
20]. The implemented method is tested on 1D NMR spectra of a Latin-square design consisting of defined mixtures of 10 different metabolites commonly found in human biofluids, as well as on real human and mouse urine spectra and human blood plasma spectra. Obtained quantitative results are compared to those of two commonly used commercial software packages, namely the AMIX software v. 3.9.13, May 2012 (
https://www.bruker.com, accessed on 2 July 2021) and the Chenomx NMR suite v. 8.6, May 2020 (
https://www.chenomx.com, accessed on 1 July 2021).