Figure 2: Description of the MetChange algorithm.

(1) Using a metabolic network reconstruction, sink (demand) reactions are added for each metabolite. Demand reactions are irreversible with the stoichiometry: metabolite −> Ø. Each demand reaction is maximized in turn to obtain maximal production values for each metabolite using a linear programming problem (LP Problem 1). (2) Reaction presence/absence P-values are generated from gene expression data and mapped onto the metabolic network. A second linear programming problem is then solved (LP Problem 2) for each metabolite. LP Problem 2 identifies the flux solution that minimizes the inconsistency of the gene expression data with the optimal production of a metabolite by restricting the demand reaction for the metabolite to be at maximal flux, and subsequently minimizing an inconsistency score of (v × P-values). (3) An example case for metabolite 1. It is observed that the control data have greater expression (lower presence/absence P-value) for certain production reactions. Greater expression of production reactions results in a lower production inconsistency score for the control gene expression sample, compared with the drug-treated case, in which certain production reactions are less expressed (higher presence/absence P-value). (4) As different metabolites have different combinations of production reactions, they cannot be compared directly within samples. Instead, scores are compared for the same metabolite between control and treated samples to generate differential consistency scores using a simple standard score. Once standardized, metabolites can be compared within drugs to identify regions where perturbation in production potential has occurred due to gene expression changes.