Automated screening for metabolites in complex mixtures using 2D COSY NMR spectroscopy

Abstract

One of the greatest challenges in metabolomics is the rapid and unambiguous identification and quantification of metabolites in a biological sample. Although one-dimensional (1D) proton nuclear magnetic resonance (NMR) spectra can be acquired rapidly, they are complicated by severe peak overlap that can significantly hinder the automated identification and quantification of metabolites. Furthermore, it is currently not reasonable to assume that NMR spectra of pure metabolites are available a priori for every metabolite in a biological sample. In this paper we develop and report on tests of methods that assist in the automatic identification of metabolites using proton two-dimensional (2D) correlation spectroscopy (COSY) NMR. Given a database of 2D COSY spectra for the metabolites of interest, our methods provide a list sorted by a heuristic likelihood of the metabolites present in a sample that has been analyzed using 2D COSY NMR. Our models attempt to correct the displacement of the peaks that can occur from one sample to the next, due to pH, temperature and matrix effects, using a statistical and chemical model. The correction of one peak can result in an implied correction of others due to spin–spin coupling. Furthermore, these displacements are not independent: they depend on the relative position of functional groups in the molecule. We report experimental results using defined mixtures of amino acids as well as real complex biological samples that demonstrate that our methods can be very effective at automatically and rapidly identifying metabolites.

Appendices

Appendix A Illustration of nomenclature

Figure A1 gives a simple example to illustrate the notation. The figure shows a portion of a map of the pixels above the detection limit for a specific metabolite superimposed on the same map for a mixture. The pixels in G only (metabolite standard) are labeled with ‘G’ and those in E only (sample) are labeled with an ‘E.’ Those pixels in both are labeled with an ‘H’, which is mnemonic for ‘hit.’ There are only three peaks for the metabolite: two on the diagonal and one off-diagonal peak that couples them. This implies two bands, one for each diagonal peak. The objective function value for b = (0,0) (i.e., no displacement) is (100) $\left({\frac{2}{3}}\right)+\frac{2}{12}$.

Figure A1.

Illustration of notation (before shift). The pixels in G only (metabolite standard) are labeled with ‘G’ and those in E only (sample) are labeled with an ‘E.’ Pixels in both are labeled with an ‘H’, which is mnemonic for ‘hit.’ There are only three peaks for the metabolite: two on the diagonal and one off-diagonal peak that couples them.

Table 5 Mixture 1, Replicate 1

Table 6 Mixture 2, Replicate 1

Figure A2 shows the result of applying a shift vector b = (1,0) to the metabolite shown in figure A1 and sumperimposed on the same sample. The resulting objective function value is given by (100) $\left({\frac{3}{3}} \right)+\frac{6}{12}$.

Figure A2.

Result of applying a shift vector b = (1,0) to the metabolite shown in Figure A1 and superimposed on the same sample.

Appendix B Detailed results of experiments

In table 5 the column labeled (P) gives the results for the full optimization model, and the columns ‘+1 pts’ ...’+4 pts’ means simple expansion of 1 pixel to 4 pixels. There are two replicates for each of the amino acid mixtures. A suitable peak detection threshold was calculated by first dividing the spectrum into 8 × 8 pixel regions, calculating the mean for each region, and then fitting the distribution of all the mean values using a Gaussian probability density function. The detection limit for the abalone data was set at 6σ, where σ corresponds to the expected standard deviation of this probability function. Note that this is a conservative limit since in analytical chemistry the limit of detection is commonly set at three times the standard deviation of the instrumental noise. For the amino acid mixtures, the estimate of the variance is artificially low because the spectra are so sparse, hence a detection limit of 100σ was visually determined to be appropriate.

Table 7 Mixture 3, Replicate 1

Table 8 Abalone digestive gland