Multivariate curve resolution of time course ¹H NMR metabolomics data

Tobias K Karakach, Mark R. Viant, Eva M. Lenz, John A. Walter

National Research Council Canada, Canada

Modeling NMR-based metabolomics data often involves linear methods such as principal component analysis (PCA), partial least square (PLS), or orthogonal signal correction.  Such methods do not generally yield factors that reflect clear biological processes, especially if the experiment is designed to investigate biological fluctuations in time, dose, or other ordinal variables.  In addition, these methods are applied under implicit assumptions about the measurement errors exhibiting an iid-normal distribution, often limiting the extent of the information recovered.

In this work, a method for the linear decomposition of NMR-based metabolomics data by multivariate curve resolution (MCR), which has been implemented elsewhere for time course transcriptomics applications [1], is introduced and implemented via alternating least squares (ALS) with a weighted least squares approach.  Measurement error information is incorporated in the modeling process allowing the least squares projections to be performed in a maximum likelihood fashion.  As a result, variability arising from pH-induced peak registration shifts can be modeled.  This eliminates the need for binning in order to control the uncertainty due to this type of variation.  The utility of the method is demonstrated using two data sets of temporal NMR metabolomics data, HgCl₂ induced nephrotoxicity in rat, and embryogenesis in a fish (Japanese medaka), which were acquired for completely different investigations.

Profiles extracted by the weighted MCR-ALS for the nephrotoxicity data exhibit strong correlations with metabolites that are consistent with temporal fluctuations in glucosuria.  For instance, in the simplest form of the model, the concentration of metabolites such as acetate, glucose, and alanine exhibit a steady increase which peaks at day 3 post-dose and gradually returns to basal levels at day 8 post-dose. On the other hand, other metabolites including citrate and 2-oxoglutarate exhibit the opposite characteristics. Although the fish embryogenesis data is more complex, the profiles extracted by the algorithm still display characteristics that depict temporal variation consistent with processes associated with embryogenesis.

[1] P.D. Wentzell, T.K. Karakach, S. Roy, M. J. Martinez, C.P. Allen and M. Werner-Washburne, BMC Bioinformatics, 2006, 7: 3 43.