Dynamic Metabolomics to Explore Models for Discovery Based Research

Joanne K. Kelleher, Department of Chemical Engineering, Massachusetts Institute of Technology
Department of Surgery, Massachusetts General Hospital

In developing the new field of metabolomics, we seek basic guidelines for the design and interpretation of models where large numbers of independent variables are measured on a small number of samples. Metabolomics shares characteristic with other -omics approaches but may have unique properties. We consider two types of metabolomics; a "static" form, where concentrations of metabolites are the measured variables, and a "dynamic" form, where isotopic tracers are used, and the steady-state enrichment of metabolites are the measured variables. For static metabolomics, the goal typically involves some aspect of classification of samples. Class comparison, class prediction, and class discovery may all be used in specific situations. Dynamic metabolomics differs in that it provides estimates of flux through a metabolic pathway.

We propose that discovery-based research may be best understood in a framework where the model for the analysis of data can be rigorously and independently tested. Dynamic metabolomics provides an opportunity for a test of this concept. We describe the analysis of dynamic metabolomic data for the gluconeogenesis pathway, in vivo, in man. This pathway may be explored with a [U- 13 C]glucose protocol (Kelleher 1999). Here plasma metabolites are sampled after steady-state infusion of [U- 13 C]glucose. The isotopic enrichment of glucose and lactate is used as input to the model to estimate gluconeogenesis. We examine discrepancies in the published models. Using simulations, we evaluated the performance of these models under various conditions.

In addition, multivariate regression models were explored with the long-term goal of developing guidelines for finding models with explanatory power. Linear models such as principal component regression and partial least-squares regression were evaluated along with a nonlinear, neural network model. We found that linear models are not strong predictors of metabolic flux given isotopic labeling data. In contrast, the nonlinear neural network model produced much more accurate estimates of flux. This may be related to the fact that the true relationship between isotope labeling and the synthesis of a polymer such as glucose (formed by condensation of two triose moieties) is not linear. We provide a basic approach to evaluate the performance of metabolomic models with the goals of classifying samples and discovering mechanisms underlying diseases. (Research supported by NIH Roadmap initiative, 1R33DK070291-01).

Kelleher, JK. 1999. “Estimating gluconeogenesis with [U- 13 C]glucose: molecular condensation requires a molecular approach.” Am J Physiol Endocrinol Metab , 277: 395-400.