Metabolomics researchers are faced with two key challenges when analyzing their NMR data. First, they must determine whether there are meaningful statistical patterns in their data. Then, they must translate these patterns into biologically relevant information in the form of metabolite identity and quantity.

In order to perform a statistical analysis of results obtained via targeted profiling in Chenomx NMR Suite, the results from a set of analyzed spectra must first be compiled into a single file. The Profiler module allows you to generate the compiled results from these analyzed spectra.

In the File menu of the Profiler module, click "Export > Compound Concentrations (.txt)..." In the resulting wizard, choose the source spectra and desired output file. For the remaining steps, export all available compounds and exclude those that have not been quantified, do not choose any extra data, and use the default export format.

The spectra for which you would like to compile analysis results must all be in the same folder. The compiled results will be stored as a simple tab-delimited text file, readable by most spreadsheet and statistical analysis software, including Umetrics SIMCA-P.

Data normalization is an important step for any statistical analysis. The objective of data normalization is to allow meaningful comparisons of samples within the dataset. When analyzing NMR spectra, dilution effects often make meaningful comparisons of data without normalization impossible; by normalizing the data you can eliminate systematic variation introduced by such effects. In the case of spectral binning (also known as "spectral bucketing"), for example, integral regions are often normalized to the total spectrum area.

To analyze targeted profiling data, two types of normalization are possible. The concentrations of all metabolites in a sample can be normalized to an endogenous metabolite that is estimated to be present at a relatively constant concentration level. Creatinine is often used for this purpose in urine studies, for example, as it is considered to be an indicator of kidney function. An alternate method for targeted profiling data is analogous to the total area normalization used in spectral binning. Metabolite concentrations can be normalized to the total concentration of all endogenous metabolites. The analysis described in this note is based on the latter form of normalization.

The desired normalization can be applied to the concentration data exported from Profiler via Microsoft Excel or any other software capable of working with tab-delimited text files.

With the concentration data exported from Profiler, and normalized as desired, you can now start the statistical analysis of the data using SIMCA-P. The version used in this application note is Umetrics SIMCA-P 11.0.

With the targeted profiling data imported into SIMCA-P, you can begin analysis of the data. The most common statistical models used with metabolomic datasets are Principal Components Analysis (PCA) and Partial Least Squares for Discriminant Analysis (PLS-DA).

One of the goals in this area of research is to use data obtained from your samples to build an appropriate model for classification and, if possible, determine the factors leading to differences among the samples. PCA and PLS-DA are often used for this purpose, as these techniques determine orthogonal latent variables that describe the input data and classify the data based on these variables. The primary advantage of using targeted profiling as an input to PCA or PLS-DA is that the resulting variables are combinations of measured metabolite concentrations. As such, these variables are easier to interpret as factors in the underlying classification model. Thus, targeted profiling provides meaningful and interpretable factors describing the input data.

PCA is a data reduction technique that may be used to reduce the dimensionality of a multi-dimensional dataset while retaining the characteristics of the dataset that contribute most to its variance.

PLS-DA is a regression extension of PCA that takes advantage of class information to attempt to maximize the separation between groups of observations.

In this dataset, you can use the available class information, specifically, the brain region that a particular sample came from, to establish a PLS-DA model. You can extract this information from the spectrum file names. First, edit the workset by clicking on "Workset > Edit > 1", and clicking on the "Observations" tab. Select "Primary ID" from the "Class from obs ID:" drop-down. Click on "Set", enter a start position of 1 and a length of 3, then click "OK". This will use the regions indicated at the start of each file name to define classes for the PLS-DA model (fcx = frontal cortex, hip = hippocampus, ocx = occipital cortex, tcx = temporal cortex). If you do not see the "Observations" tab, you may need to switch to Advanced Mode by clicking the appropriate button in the bottom left corner of the workset wizard.

Once the data is filtered and class values are identified, you can generate a PLS-DA model. Click "Analysis > Change Model Type" and select "PLS Discriminant Analysis". To fit the model to the data, select "Analysis > Autofit", which will extract the principal components needed to properly fit the model.

Having determined that there is a clear pattern in the data, it is natural to ask which metabolites are responsible for separating the various elements of the brain. In other words, what is the biological significance of the data? The loadings plot can provide an answer to this question, by describing the weighting coefficients for each metabolite.

Chenomx would like to thank Brent McGrath and Dr. P. Silverstone in the Department of Psychiatry at the University of Alberta for providing data for this application note.