All known metabolic reactions were considered as candidate heterologous reactions that could be added to the host metabolic network. We first constructed an in-house database of metabolic reactions from data stored in the KEGG ligand section 16 and BRENDA 17 databases. All metabolic reaction information regarding genes, enzymes, pathways, and organisms in the KEGG database was collected into the database, which was developed using PostgreSQL 9.0 (The PostgreSQL Global Development Group). The Michaelis-Menten constants (K _m ) of the enzymatic reaction data were retrieved from BRENDA 17 . We also used Python scripts to access the in-house database.

In this study, we adopted 3 host microorganisms widely used in industry; namely, E. coli C. glutamicum, and S. cerevisiae. E. coli has been exploited for such industrially valuable compounds as L-phenylalanine, L-tyrosine, 1-butanol and 1,2-propanediol 32 33 34 . C. glutamicum is widely used in amino acid production 35 . S. cerevisiae is an important producer of alcohols and organic acids such as lactate 36 . These organisms are ideal hosts of bioengineered products since they exhibit high growth activity under various conditions and are easily genetically manipulated 37 38 .

We used genome-scale metabolic models of S. cerevisiae (iMM904) 39 , E. coli (iJR904) 40 , and C. glutamicum 41 , based on earlier metabolic constructions with slight modifications. Because our pathway search algorithm uses the heterologous reactions listed in the KEGG database, all metabolite IDs in the earlier genome-scale metabolic models were converted to the KEGG compound ID format using metabolite name matching by manual checking.

We developed an algorithm to identify heterologous reaction(s) producing a target metabolite within a host microorganism. The algorithm expands the host metabolic network by sequentially adding heterologous metabolic reactions from our in-house database. The heterologous pathway identification procedure is as follows:

1. A set of metabolites M0 and a set of metabolic reactions R0 are defined as those present in the genome-scale metabolic network of the host microorganism.

2. From the in-house database, heterologous reactions that satisfy the following conditions are collected: (i) the reaction does not exist in R0, and (ii) it can produce metabolites that do not exist in M0 from a metabolite in M0. A set of these heterologous reactions is defined as R1, and a set of metabolites produced by reactions in R1 is defined as M1.

3. In the same way, Ri is the set of reactions not present in {R0, R1, … , Ri − 1} which can produce metabolites not existing in {M0, M1, … , Mi − 1} from metabolites included in those sets. This expansion procedure is iterated until no further reaction is connectable to the expanded metabolic network.

If a target metabolite is included in a nonnative metabolite set M _i , we can identify a set of heterologous reactions that are necessary to produce the target metabolite. For simplicity, all metabolic reactions in the database were assumed to be reversible. Of course some reactions are known to be irreversible, such as the carboxylation and decarboxylation reactions classified by Nomenclature Committee of the international Union of Biochemistry and Molecular Biology (NC-IUBMB) 42 . However, for the majority of reactions in the database, directional information is limited and thus the reversibility of the reactions is difficult to judge. By assuming that all reactions are reversible, we avoid the risk of missing important heterologous pathways due to misjudgment of their reaction reversibility. Our strategy here is to initially screen all possible heterologous pathways regardless of reaction irreversibility, then decide whether the predicted pathway is plausible based on physiological knowledge of the reaction irreversibility.

FBA is based on a genome-scale metabolic model and optimization of a specific objective flux by linear programming 43 44 . We used FBA to estimate the metabolic flux profile of metabolic networks expanded with heterologous reactions. A pseudo-steady state is assumed, that is, the net sum of all production and consumption fluxes for each internal metabolite is zero. In matrix notation, this condition is represented as S ⋅ v = 0 _, where S is the stoichiometric matrix representing the stoichiometry of metabolic reactions in the network and v is the vector of metabolic fluxes. In FBA, the flux profile (constrained by steady state) is determined by optimizing a specific objective function. The biomass production flux is one of several widely used objective functions that can be maximized. The flux profiles obtained by maximizing biomass production fluxes are known to be well correlated with those obtained experimentally 39 40 41 45 .

In this study, the coefficients of metabolites representing biomass production flux were extracted from earlier studies 39 40 41 . We employed another objective function, the production flux of the target metabolite, to judge whether the target metabolite was producible by the metabolic network. In all of the FBA simulations in this paper, glucose was chosen as the sole carbon source and the following external metabolites were allowed to freely transport through the cell membrane: CO₂, H₂O, SO₄ or SO₃, and NH₃. All calculations were performed using MATLAB 2009b (MathWorks Inc., Natick, MA). The linear programming problem was solved using GLPK 4.34 (GNU Linear Programming Kit) 46 via the MATLAB interface.

7,769 metabolic reactions and 6,635 metabolites (shown in the Additional file 1) from 1,139 species were collected from the KEGG database and deposited in our in-house database. To screen for target metabolites that could be produced by our host microorganisms S. cerevisiae, E. coli, and C. glutamicum, we iteratively expanded the host metabolic network by adding heterologous metabolic reactions as described in the Methods section. Figure 1 displays the number of nonnative metabolites connected to the host metabolic network as a function of the number of heterologous reactions. Fewer than 33 heterologous reactions are required to connect 3,154, 3,244, and 3,112 nonnative metabolites to the host metabolic networks of S. cerevisiae, E. coli, and C. glutamicum respectively.

The list of metabolites connected to the host metabolic networks is presented in the 2 3 4. To this list, we added the K _m values of heterologous enzymes. Knowing the K _m assists in deciding which heterologous enzymes originating from various organisms should be introduced to the host. The names of organisms in the BRENDA database displaying minimum K _m of the corresponding heterologous enzymes are also listed 17 , since the enzyme from this organism is expected to have highest affinity among the orthologous enzymes to the corresponding substrate. Importantly, these identified heterologous reactions of nonnative metabolite production agreed well with those widely used in metabolic engineering and which are important to the industry (Table 1), such as isoprene, α-farnesene, poly-β-hydroxybutyrate (PHB), and cadaverine.

As an example, the production pathways of 1,3-propanediol (C02457) by E. coli and S. cerevisiae, which were adopted in earlier studies 52 53 , are shown in Figure 2. In the previous studies, C02457 production proceeded via conversion of glycerol to 3-hydroxypropanal using glycerol dehydratase (encoded by dhaB1-B3). 1,3-Propanediol was then produced, aided by 1,3-propanediol oxidoreductase (encoded by dhaT). In this study, the screened heterologous pathways for C02457 production exactly matched those of the earlier studies. In E. coli, the screened production pathways of isoprene, α-farnesene, and PHB derived by our algorithm were also identical to those of the earlier studies, while similar heterologous genes introduced to the alternative hosts (C. glutamicum and S. cerevisiae) additionally produced these targets (see Table 1). Moreover, both reported and alternative production pathways were screened by our algorithm. For instance, we found that E. coli cells can produce (R)-propane-1,2-diol when methylglyoxal reductase and lactaldehyde reductase are added to the metabolic network, which has not been reported to date. Similar alternative pathways were found for the production of itaconate, cis cis-muconate, and 2,3-dihydroxybenzoate. These results suggest that our algorithm successfully identified the metabolic reactions necessary for the target productions and could assist in screening for potential host cells.

Next, we used glucose as a carbon source to investigate whether these nonnative metabolites are producible by FBA simulations. In this simulation, the production flux of each nonnative metabolite was treated as an objective function to be maximized under the steady-state assumption. When the maximum production flux of a nonnative metabolite is zero, this metabolite is non-producible under the given condition.

We calculated the maximum production fluxes of all connectable nonnative metabolites. 28% of the connectable nonnative metabolites of E. coli could not be produced using glucose as a sole carbon source. Similarly, 33% of the connectable nonnative metabolites of S. cerevisiae and 16% of the connectable nonnative metabolites of C. glutamicum were non-producible under this condition. These non-producible metabolites were identified by their tendency to disconnect when glycolysis formed the central metabolic pathway. In E. coli, these metabolites included trans-aconitate (C02341), butyrate (C00246), acetoacetate (C00164), and l-lactaldehyde (C00424).

To evaluate the feasibility of nonnative target metabolite production, we performed FBA simulations under conditions of maximizing biomass production following heterologous reaction expansion of the genome-scale metabolic model. Metabolic flux profiles calculated at maximum biomass production rates have been shown to closely represent those in real microorganisms 45 59 60 61 62 . Such agreement may be explained by the growth optimization of microorganisms through evolutionary dynamics 63 . Furthermore, for the mutant strains constructed in the laboratory, the cells could achieve the near-optimal metabolic state calculated by the FBA simulation after long-term cultivation 64 65 66 67 , via the selection of faster growing cells. Thus, we can expect that if a nonnative target metabolite is produced in the FBA simulation under maximized biomass production, that target may be feasibly manufactured.

In Figure 3, we plot the number of target metabolites produced under maximized biomass production, versus the number of heterologous reactions necessary for metabolite production. We set a threshold yield (1%) to identify the produced metabolites because the production yields of some metabolites were positive but extremely small. Sometimes the FBA solution was undetermined under biomass maximization conditions; that is, the solution was not unique. In such cases, following maximization of biomass production, the production flux of the target metabolites was further maximized with fixing the maximized biomass production, to obtain a unique flux profile that would generate the target. In the simulations, we adopted a micro-aerobic condition to screen the target metabolites produced under the biomass maximization condition, in which significantly more metabolites were obtained than under anaerobic conditions, and in which all anaerobically produced metabolites were included.

Table 2 lists the representative target metabolites produced under biomass maximization, together with their corresponding heterologous reactions. The mechanisms involved in these reactions can be classified into two categories. One is based on the production of oxygen as a by-product of the targets. Since the simulations were performed under micro-aerobic conditions, oxygen supply increased the biomass production by activating the electron transfer system and facilitating adenosine triphosphate production. Therefore, if the heterologous reactions used to produce the target are accompanied by oxygen production, the target can be produced under minimum biomass production flux. For example, pentane-2,4-dione was produced by introducing a single heterologous reaction into E. coli and S. cerevisiae, whereas two heterologous reactions were necessary to produce this metabolite in C. glutamicum. Vanillin can be produced under the same mechanism by introducing 4 heterologous reactions into the E. coli and C. glutamicum metabolic networks.

Another mechanism is associated with NADH oxidization. Under micro-aerobic conditions, the cellular growth of microorganisms can be limited by NAD regeneration, which is necessary for glycolysis activity, and which occurs through NADH oxidization. Thus, when the heterologous reactions producing the targets are associated with NADH oxidization, these heterologous reactions are activated when the biomass production is maximized This phenomenon occurs, for example, in the production of (R)-propane-1,2-diol and 2-propyn-1-al.

We also found that some metabolites are produced only by E. coli under conditions of maximum biomass production, such as (R)-propane-1,2-diol and adipate semialdehyde. Unlike S. cerevisiae and C. glutamicum, E. coli possesses NAD transhydrogenase, which can convert NADP and NADH to NADPH and NAD  respectively (and vice versa). In E. coli cells, the excess NADH is converted to NADPH which can then enter the target production pathway.

While screening for heterologous pathways to produce the target metabolites discussed earlier, differences in production capacity between the three host microorganisms emerged; for example, a group of metabolites was inducible by the addition of heterologous reactions to one of the hosts, but was not produced by the other hosts. To characterize the differences in target production capacity, we categorized the producible metabolites (shown in the Additional files 5 6 7) using the KEGG Orthology database 16 . We then performed a chi-square statistical analysis to identify the categories in which the frequency of producible metabolites is significantly higher than expected. Figure 4 shows the 10 categories that demonstrated significant differences (P < 0.001). As shown in the figure, metabolites belonging to 5 categories, namely, “tyrosine metabolism,” “dioxin degradation,” “benzoate degradation,” “chlorocyclohexane and chlorobenzene degradation,” and “xylene degradation,” tended to be producible by S. cerevisiae and C. glutamicum but were scarce in E. coli cells.

Similarly, the metabolites in “flavonoid biosynthesis,” “phenylpropanoid biosynthesis,” and “nicotinate and nicotinamide metabolism” were preferentially generated by E. coli and C. glutamicum. Metabolites assigned to “porphyrin and chlorophyll metabolism” also tended to be produced in C. glutamicum cells. Likewise, the metabolites assigned to “biosynthesis of 12-, 14-, and 16-membered macrolides” were produced preferentially in E. coli cells. Such differences in production capabilities result from the different metabolic pathways by which the hosts produce necessary substrates, and from cellular compartmentalization in the yeast strain (which is absent in the bacterial strains).

In yeast cells, the compartments present barriers to metabolite transport. For instance, mitochondrial/cytoplasmic interfaces prohibit the production of certain target metabolites when sugar is used as a carbon source. Similarly, the production of metabolites in the “flavonoid biosynthesis” category was inhibited in yeast cells because the transportation of 4-coumarate between the mitochondria and the cytosol is not permitted; therefore, the yeast strain could not produce p-coumaroyl-CoA (required for making chalconoid, an important ingredient in flavonoid biosynthesis). Our genome-scale metabolic model does not account for transportation capabilities between compartments, which are currently unclear for many metabolites, and which might influence the production capacities of target metabolites in real cell systems.