Globalfit: Refining Metabolic Networks By Simultaneously Matching Growth and Non-Growth Data Sets

Genome-scale metabolic models are refined through comparisons to experimental growth data from gene knockouts or nutrient environments. We present a novel method that identifies the minimal set of network changes needed to correctly predict all experimentally observed growth and non-growth cases simultaneously. Applied to the genome-scale model for Mycoplasma genitalium, GlobalFit decreased unexplained cases by 79%, boosting accuracy from 87.3% (according to GrowMatch [1], the current state-of-the-art) up to 97.4%.

All currently available methods for network refinement based on growth data are greedy algorithms, solving one inconsistency between model and experiment at a time. While each individual set of network changes is minimal, the union of these sets can become unrealistically large. Model changes introduced early may make the reconciliation of false predictions considered later impossible, and experimental errors that happen to be consistent with the initial model can severely bias the results

GlobalFit is a novel bi-level optimization method capable of comparing flux-balance analysis (FBA) model predictions to measured growth across all tested environments and gene knock-outs simultaneously. Allowed model changes are (i) removals or (ii) reversibility changes of existing reactions; (iii) additions of reactions to the model from a database of potential reactions; (iv) removals of metabolites from the biomass; and (v) additions of metabolites to the biomass.

The algorithm is first formulated as a bi-level linear problem, where each condition is represented by separate metabolites and fluxes. To ensure in silico growth for conditions with experimentally demonstrated growth, the biomass production for these conditions must be greater than a predefined threshold. For non-growth phenotypes, the inner optimization problem maximizes the biomass production to check whether it stays below a non-growth threshold. The outer optimization problem jointly minimizes the number of model changes and the number of experiments that are incorrectly predicted by the final model. The penalties for individual network changes can be set independently. This allows, for example, to prefer reversibility changes over reaction additions, or to preferentially include reactions from metabolic network reconstructions of close relatives. The bi-level problem can be re-formulated as a single-level optimization problem; a corresponding implementation of GlobalFit, integrated with the sybil toolbox for constraint-based analyses [2], is freely available from CRAN (http://cran.r-project.org/web/packages/GlobalFit/).

1.         Satish Kumar, V. and C.D. Maranas, GrowMatch: an automated method for reconciling in silico/in vivo growth predictions. PLoS Comput Biol, 2009. 5:e1000308.

2.         Gelius-Dietrich, G., et al., Sybil--efficient constraint-based modelling in R. BMC Syst Biol, 2013. 7:125.