Figure 3. Using models for qualitative predictions and iterative improvement.

A. Each reaction in the network is linked to a protein and encoding gene through the gene-protein-reaction (GPR) relationship. Because each reaction in the network corresponds to a column in the stoichiometric matrix, simply removing the column association with a particular reaction can simulate gene knockouts. Thus, multiple KO simulations can be performed. For example, it is easy to delete every pairwise combination of 136 central carbon metabolic E. coli genes to find double gene knockouts that are essential for survival of the bacteria.

B. The simplicity of altering inputs to change cellular growth environments and removing genes in silico allows one to perform simulations in millions of experimental conditions quickly. Even on a modest laptop computer a single FBA calculation runs in a fraction of a second, thus simulating the effect of all gene knockouts in E. coli central metabolism can be run in less than 10 seconds.

C. Incorrect model predictions are an opportunity for biological discovery because they highlight where knowledge is missing. Targeted experiments can be performed to discover new content that can then be added back to a model to improve its predictive accuracy. Missing model content can be discovered using automated approaches known as ‘gap-filling’ (Orth and Palsson, 2010a) that query a universal database of potential reactions to restore in silico growth to a model.

D. Gap-filling approaches have been used to discover new metabolic reactions in several organisms. E. coli: Two new functions for two classical glycolytic enzymes phosphofructokinase (PFK) and fructose-bisphosphate aldolase (FBA) were discovered (red) (Nakahigashi et al., 2009a). Human: Gluconokinase (EC 2.7.1.12) activity was discovered based on the known presence of the metabolite 6-phosphogluconolactonate in the human reconstruction (Rolfsson et al., 2011b) (red). Yeast: Automated model refinement suggested modifications in the NAD biosynthesis pathway. Experiments demonstrated that a parallel pathway from aspartate thought to exist in yeast was not present (Szappanos et al., 2011).

E. False positive predictions can be reconciled by adding regulatory rules derived from high throughput data (Covert et al., 2004), for example, a recent study was able to reconcile 2,442 false model predictions from the E. coli GEM by updating the function of just 12 genes (Barua et al., 2010). Additionally, a false positive growth inconsistency in the metabolic model of S. Typhimurium was reconciled by updating regulatory rules for the iclR gene product’s transcriptional repression of aceA encoding isocitrate lyase. Transcriptional repression can also often be relieved via adaptive laboratory evolution. Such evolution drives experimental phenotypes to achieve model predictions. Several experimental studies have shown that an organism can evolve to achieve model-predicted optimal growth state (Ibarra et al., 2002).