MetaMerge: scaling up genome-scale metabolic reconstructions with application to Mycobacterium tuberculosis

Abstract

Reconstructed models of metabolic networks are widely used for studying metabolism in various organisms. Many different reconstructions of the same organism often exist concurrently, forcing researchers to choose one of them at the exclusion of the others. We describe MetaMerge, an algorithm for semi-automatically reconciling a pair of existing metabolic network reconstructions into a single metabolic network model. We use MetaMerge to combine two published metabolic networks for Mycobacterium tuberculosis into a single network, which allows many reactions that could not be active in the individual models to become active, and predicts essential genes with a higher positive predictive value.

Figure 1

Flow chart for the MetaMerge algorithm. Shown is a flow chart of the six stages of the MetaMerge algorithm. Stage 1: the two models to be merged are parsed and the features for metabolites and reactions are prepared. Stage 2: an initial matching of reactions and metabolites is created. Stage 3: the user is asked to confirm the newly matched reactions and metabolites (optional step). Stage 4: the newly matched metabolites are used to extend the current matching of reactions and metabolites. Stages 3 and 4 are repeated until convergence. Stage 5: the matching of metabolites is checked to ensure transitivity and non-transformability. Stage 6: the merger is performed and the resulting network is output in the desired format. Circles indicate metabolites (red, model 1; white, model 2; pink, combined); squares indicate reactions (turquoise, model 1; brown, model 2; dark green, combined). The mini-networks were created using the Cytoscape software [34], version 2.7.0.

Figure 2

The combined metabolic network for M. tuberculosis. (a) The combined model. Shown is the combined model for M. tuberculosis generated by MetaMerge. The network was laid out with Cytoscape [34] and 12 isolated reactions were removed from the final figure. The color scheme is identical to that of Figure 1, with the Beste et al. model [13] used as model 1 and the Jamshidi and Palsson model [14] used as model 2. (b) Example pathways in the original and combined models. Shown are the pentose phosphate and glyoxylate metabolism pathways in the original models (model 1, labeled GB [13], and model 2, labeled JP [14]) and the combined model. The enzymes catalyzing each reaction are included as a top layer, and their names are shortened by removing Rv. Circles indicate metabolites (red, model 1; white, model 2; pink, combined); squares indicate reactions (turquoise, model 1; brown, model 2; dark green, combined); octagons indicate enzymes (light green, model 1; yellow, model 2; magenta, combined). The subnetworks were laid out with Cytoscape [34] and several currency metabolites were removed for visual clarity.

Figure 3

Essentiality predictions in model I and in the combined model. Shown are receiver operating characteristic curves for the correctly predicted fraction of gene essentiality (y-axis) based on model 1 and the combined model when the essentiality threshold for the TraSH experiment [19] is allowed to vary between 0 and 0.2 in increments of 0.002 (x-axis).

Figure 4

Enzymes predicted to have similar metabolic impact to that of isoniazid targets. Shown are the enzymes in the combined model identified to be in an enzyme subset with the targets of isoniazid, as well as the reactions for which these enzymes are essential. Circles indicate metabolites; squares indicate reactions; octagons indicate enzymes. The subnetwork was laid out with Cytoscape [34] and several currency metabolites were removed.