Schematic procedure of the framework for network modularization and Bayesian network analysis (FMB). (A and B) Metabolic network model is repeatedly simulated for control and perturbation conditions using constraints-based flux analysis with constraints from 13C-based metabolic flux analysis and cell culture to calculate more reliable metabolic flux distributions while amplifying the data size. (C) The result is a flux matrix (N-by-M) that contains a total of 2,000 samples for each reaction from both control and perturbed condition. (D) Core reactions (see main text for definition) are selected in this step. (E) Flux matrix is converted to flux-pattern matrices that contain information on flux variation pattern from sample to sample, having one of ‘1’, ‘-1’ and ‘0’ (see Methods for details). All the generated flux-pattern matrices are adjoined into a single large flux-pattern matrix for clustering. (F) Hierarchical clustering is applied to this matrix, and reactions are clustered in terms of the uniform functionality, creating metabolic clusters. (G) Bayesian network (BN) of each metabolic module is first inferred, producing local scale BNs, and representative reactions, the most influential ones in their corresponding metabolic module, are determined by measuring the degree of influence of each reaction on others in a module using total mutual information (TMI). Mutual information (MI) is calculated between a target node (blue color) and the other remaining nodes pair by pair, indicated as MI1, MI2, MI3 and MI4, in a local scale BN of metabolic module. TMI of the reaction is a summation of these MI values. This procedure is repeated until TMIs of all the other reactions are calculated. At the end, reaction with the highest value of TMI is selected as the representative reaction of the metabolic module. (H) Representative reactions are finally subjected to BN analysis to infer a global scale BN for detailed analysis of specific perturbation given to the biological system.