MetaboLab - advanced NMR data processing and analysis for metabolomics

One-dimensional NMR

Metabolomics has become an important technique in the context of systems biology to characterize changes in metabolite composition and concentration in biological systems such as cells, tissues or in bio-fluids. One-dimensional (1D) NMR spectra used in the context of metabolomics contain hundreds of signals arising from 50-100 metabolites. To utilize this wealth of information in the context of statistical analysis, consistent and accurate processing of the data is paramount. This includes phase correction of complex NMR signals to pure absorption line shapes and consistent baseline correction across series of spectra, along with various linear and non-linear scaling algorithms and spectral alignment (see additional file for more information). Scaling includes linear scaling algorithms, specifically total spectral area scaling and probabilistic quotient normalization [1]. Among non-linear scaling algorithms it includes the generalized logarithmic transformation (glog), Pareto or autoscaling [2], used prior to statistical processing. Excellent reproducibility within series of spectra is essential for subsequent statistical analysis using multivariate algorithms such as principal component analysis (PCA) or PLS-DA, but also for univariate analysis and signal integration.

To become accessible to a broader range of users in a translational setting, metabolomics software needs to provide intuitive and transparent control over all processing steps, without limiting more sophisticated uses. This need has been addressed by a batch processing interface, suitable to handle larger series of spectra with standard processing parameters, with an option to create user editable scripts allowing more sophisticated changes and providing access to algorithms from other packages.

Two-dimensional NMR spectra

The fundamental requirements for spectral processing of two-dimensional (2D) NMR spectra, such as J-resolved or HSQC spectra, are similar to those for 1D-NMR spectra. Owing to the reduced congestion in 2D spectra and the smaller number of data points, algorithms such as baseline correction are easier to implement. 2D spectra also present additional opportunities, especially for the assignment of metabolites depending on user driven decisions for ambiguous signals. This requires a spectral library, ideally one that has been recorded for the same field strength and sample conditions (e.g. solvent composition, buffer, sample pH and temperature). MetaboLab implements an easy mechanism to add spectral data for metabolites taken from databases.

Assignment decisions will often depend on the size of signal intensities in relation to other signals of the same metabolite, facilitated by the display of intensity values for picked peaks within regions of interests for the resonances of the metabolite of interest. MetaboLab also includes a simple mechanism to transfer peak lists between spectra and automatically adjusts small shifts of signals arising from pH variations.

Objectives

One-dimensional NMR spectra processing and batch processing

MetaboLab includes the processing of the original spectrometer data (implemented for Bruker and Varian/Agilent data and parameter files), followed by common processing steps. The free induction decay (FID) is Fourier transformed, phase corrected and referenced, a robust automated algorithm is included for phase correction, which has worked well for most spectra acquired with the Bruker 'baseopt' acquisition parameter. Automated referencing to chemical shift standards such as TMS, TMSP or DSS is available. If no chemical shift standard has been added referencing is performed on the assumption that the transmitter offset has been set to the water frequency [8].

For batch processing of NMR series of NMR spectra a script builder generates scripts using standard processing steps including zero filling, several algorithms for post acquisition water suppression [9], apodization functions, phase correction (two different automated algorithms or manual phase correction), DC offset correction, Gibbs multiplication and referencing. The file selection mechanism allows the simultaneous choice of NMR data from different directories. The script builder composes a MATLAB script, which can be further edited and executed. This script will read all the selected FIDs and process them subsequently (see also Additional File 1).

Processing scripts can similarly be composed for 2D spectra. Parameter sets for 2D J-resolved and for HSQC spectra have been included. For a series of 2D J-resolved spectra a projection using either the skyline or summation algorithm along the frequency axis can be applied after optionally calculating a tilted 2D spectrum. Apodization functions default to either exponential multiplication for 1D spectra or square cosine for 2D spectra. A window function composed of a shifted sine multiplied by an exponential (SEM window function [10]) is also available for 2D J-resolved spectra.

MetaboLab: Metabolomics post-processing

Metabolomics NMR data analysis requires additional post-processing applied to the data series in order to remove distortions causing artifacts in subsequent statistical data analysis or quantification. Post-processing includes spectral alignment (with respect to TMSP or a reference spectrum, the latter is typically performed after selected regions of the spectra have been excluded), various algorithms for baseline correction, the scaling of spectra (with respect to TMSP or total spectral area), bucketing and data export for subsequent multivariate statistical analysis. The graphical user interface (Figure 1) allows for the exclusion of spectra and the assignment of classes, which can be displayed in different colors. An interface is provided to transfer data and spectral classes into the commercially available PLS Toolbox software (Eigenvector Research Inc., Wenatchee, USA).

The MetaboLab graphical user interface provides easy access to post-processing tasks. Even on modern spectrometers the spectral baseline is often not ideally flat or shows a small DC offset. Such distortions typically may arise from the use of digital filters in the context of oversampling or from imperfections in the water suppression. Moreover, small baseline distortions in high-resolution 1D spectra are enhanced by non-linear scaling algorithms such as the glog transform [2], Pareto or autoscaling, which are typically used to enhance small signals in spectra. MetaboLab includes various algorithms with automated determination of baseline points for baseline correction [11, 12]. However, parametrization of these algorithms is often difficult and tends to cause artifacts at the edge of peaks. Best results are achieved by manually picking baseline points on the series of NMR spectra, using common zeros in all or most spectra, followed by calculating a spline function to model the baseline for each individual spectrum. The advantage of this is that exactly the same spline points can be used for all spectra, thus reducing the variability between spectra. Facilitated by a graphical interface this helps to minimize distortions introduced by the baseline correction.

MetaboLab includes a new algorithm to compute an optimized glog parameter that maximizes the variance in the spectral series of NMR spectra. For sufficiently large data sets this algorithm yields good glog parameters. MetaboLab also allows manual modification of glog parameters and facilitates its adjustment by updating the display of processed spectra in real time.

Graphical Assignment Tool

The assignment tool presents itself as a graphical user interface to assign signals in 2D-HSQC spectra based on metabolite peak lists in a spectral library (Figure 2). MetaboLab contains a limited number of assigned metabolites from public databases [13] but provides an easy method to add additional metabolites. The software allows for small differences in chemical shifts in a spectrum compared to those in the library and searches for the closest local maximum (similar to the region of interest based data analysis in the rNMR program [14]). The chemical structure including atom numbering of the carbon nuclei is displayed for an individual metabolite (see Additional File 1 for details). For ambiguous assignments alternative assignments available in the spectral library can be displayed along with peak intensities. Duplicate assignments are highlighted by color.

Automated phase correction

The automatic phase correction algorithm uses the TMSP signal if present, employing baseline properties between the TMSP signal and its ¹³C-satellites. It also uses baseline properties at the other end of the spectrum to calculate a first order phase correction. If no TMSP is present it uses baseline properties at both ends of the spectrum to determine a zero order phase correction. Baseline correction. The data structure used in MetaboLab reflects closely that of NMRLab [3]. It uses a MATLAB data structure (metaboSpc) with fields storing the information for the different data post-processing steps, e.g. metaboSpc.baseline contains further data structure fields for the different baseline correction algorithms. For example, metaboSpc.baseline.spline contains the field baseline_points as a vector, the field baseline_average_points as a number indicating how many baseline points are averaged around each baseline point (selected in baseline points), the field linear determining whether a linear interpolation of adjacent baseline points is to be used for regions where no spectral baseline is available within a range of linear_points. All post-processing steps can be reconstructed from the information stored in the metaboSpc data structure.

Samples

1D-¹H NMR spectra shown in Figures 1 were acquired from ultra-filtrated blood plasma samples [15]. NOESY-presat was used to suppress the solvent resonance. All spectra were automatically processed, phase corrected and referenced using the script builder interface before data post-processing was performed using the Metabolab GUI software. 2D-HSQC spectra were acquired from MeOH/CHCl₃ cell extracts of K562 CML cells fed with ¹³C(1,2)-labeled Glucose, as described in [16] except for the addition of labelled glucose. NMR spectra. HSQC spectra were recorded using echo/anti-echo coherence selection for quadrature detection in the indirect dimension. The first spectrum was manually processed using the NMRLab software, then all spectra were automatically processed with the script builder interface using referencing and phasing information from the first spectrum. Resonance assignment was performed inside the HSQC assignment tool.