Modelling and simulation for metabolomics data analysis

The advent of large data sets, such as those produced in metabolomics, presents a considerable challenge in terms of their interpretation. Several mathematical and statistical methods have been proposed to analyse these data, and new ones continue to appear. However, these methods often disagree in their analyses, and their results are hard to interpret. A major contributing factor for the difficulties in interpreting these data lies in the data analysis methods themselves, which have not been thoroughly studied under controlled conditions. We have been producing synthetic data sets by simulation of realistic biochemical network models with the purpose of comparing data analysis methods. Because we have full knowledge of the underlying ‘biochemistry’ of these models, we are better able to judge how well the analyses reflect true knowledge about the system. Another advantage is that the level of noise in these data is under our control and this allows for studying how the inferences are degraded by noise. Using such a framework, we have studied the extent to which correlation analysis of metabolomics data sets is capable of recovering features of the biochemical system. We were able to identify four major metabolic regulatory configurations that result in strong metabolite correlations. This example demonstrates the utility of biochemical simulation in the analysis of metabolomics data.