MicrobesFlux can serve as a platform to build metabolic models for 1,194 organisms sequenced in the KEGG database. To generate a metabolic model for a specific microorganism, the gene annotation information is downloaded from a set of corresponding KGML files, and cross-referenced with the KEGG LIGAND database. The generated pathway network serves as a seed model subject to further customization by the users.

The model draft of the designated microorganisms, automatically generated from the KEGG database, should be reconstructed to fill in the gaps between the genome annotation and the functional metabolism. MicrobesFlux can provide a list of tools to help users build customized metabolic models. They include: 1) the introduction of inflow/outflow fluxes to describe the nutrient transport between the cytosol and the extracellular medium; 2) the implementation of biomass production reactions with the manual identification of the organism-specific biomass composition; 3) the inclusion of heterologous pathways or gap-fill pathways that do not exist in the seed model; and 4) the modification of metabolic pathways by tuning reaction directions and reactants/products species in the reactions. The functions in MicrobesFlux can help users build a drafted metabolic model for further reconstruction and flux analysis. The metabolic pathway network can be exported as SBML and viewed with SVG.

It is necessary to point out that the complete genome-scale metabolic model reconstructions cannot be automatically achieved in MicrobesFlux (Table 1) yet. According to the protocol made by Thiele and Palsson 16, a high-quality genome-scale metabolic reconstruction requires 4 stages and 94 steps. By using MicrobesFlux, users can finish 100% of stage 1 (i.e., draft recognition, 5 out of 5 steps), 69% of stage 2 (i.e., refinement of reconstruction, 22 out of 32 steps), 100% of stage 3 (i.e., conversion of reconstruction into a computable format, 5 out of 5 steps) and 33% of stage 4 (i.e., network evaluations, 17 out of 52 steps). More specifically, several iterative model curation steps in stage 2 and 4 cannot be achieved in the current version of MicrobesFlux yet. For example, the compound charging (step 8 in stage 2) and charge balance in metabolic reactions such as electron transport chains (step 20 in stage 2) cannot be automatically curated due to the lack of information in the KEGG LIGAND database and KGML files used for building the metabolic models in MicrobesFlux. In addition, the automated computational gap-analyses (step 46–48 in stage 4) and gap-fill algorithms (step 51–58 in stage 4) are not incorporated in MicrobesFlux, although, for a medium-scale metabolic model, users can manually identify gaps (such as dead-end metabolites) and adjust the metabolic reactions to fill these gaps. Other features that are essential for completing a high-quality genome-scale metabolic model but are not covered by MicrobesFlux include predicting reaction directions based on thermodynamics (step 10), assigning the location of genes and metabolic reactions (step 11), and correlating genes, proteins and metabolic reactions (step 13). In order to construct an accurate genome-scale metabolic model, these steps need to be done by users off-line to expand the MicrobesFlux-drafted model. Given this, the MicrobesFlux software should be used as a toolkit for “drafting” instead of “finalizing” metabolic models.

For each reconstructed metabolic network, users can choose to perform either FBA or dFBA tasks to determine flux distributions under metabolic steady or dynamic states. The objective functions used in constraint-based flux analysis can either be the commonly used “maximizing biomass”, or a user-defined objective function. The users can also define the upper and lower bounds for each flux in FBA and dFBA. MicrobesFlux converts the constraint-based flux analysis model on the reconstructed network to a large-scale optimization problem with constraints, and employs IPOPT (an optimization solver based on an interior-point algorithm) to find an optimal solution 24. The solutions acquired by IPOPT are then transformed into the metabolic fluxes as requested by users. MicrobesFlux automatically distributes these large-scale optimization tasks to a cluster of machines so that multiple users can conduct flux analyses on our system simultaneously. Our system also supports the dynamic simulation of metabolic fluxes by using the static optimization approach (SOA) 25. In this approach, the users can assume that the entire dynamic microbial metabolism is composed of numerous pseudo-steady states. For each pseudo-steady state, a conventional FBA problem is formulated with the user-defined inflow and outflow fluxes. In other words, a dFBA problem will be converted to multiple mini-FBAs that are subject to constraints from the measurement of time-dependent inflow (substrate uptake) and outflow fluxes (metabolite production). To avoid extensive analytical efforts to measure inflow and outflow fluxes at each time interval, the users can use an empirical or a kinetic model to estimate the time-dependent inflow/outflow fluxes for the mini-FBAs through an entire growth period based on limited measurement data. The dynamic flux simulation is of particular industrial interest 26, since many biological systems cannot maintain a meaningful metabolic steady state during the fermentation process.

To demonstrate the use of the MicrobesFlux platform, a simple toy model was constructed, which only included the central metabolic pathways: the glycolysis pathway, the pentose phosphate pathway, the TCA cycle, and the anaplerotic pathway. Glucose represented the carbon substrate and acetate represented the extracellular metabolite product. The TOY model was loaded from MicrobesFlux (Figure 3), which included 10 reactions that described the intracellular fluxes and lumped biomass production. Subsequently, the toy model was reconstructed by introducing the inflow flux: “Glucose → G6P” and the outflow flux: “AcCoA → Acetate”. The drafted TOY model was then used for constraint-based flux analysis by setting the objective function as “maximizing biomass” and fixing the inflow and outflow fluxes as 11.0 and 6.4 mmol/g DCW/h respectively. The reconstructed pathways and the simulated flux distributions from the TOY model can be found in Additional file 1 and Additional file 2. The simulated results from MicrobesFlux were confirmed by an independent linear optimization of the same TOY model via “linprog” in MATLAB (Additional file 3).

Based on a similar flowchart to that of Case Study 1, we drafted a medium-scale metabolic model of Thermoanaerobacter sp. strain X514. Thermoanaerobacter species are thermophilic bacteria that can simultaneously convert both pentose and hexose to ethanol with high yield 28 29. It can also be co-cultured with cellulosic Clostridium thermocellum species to produce ethanol from cellulose. Thermoanaerobacter sp. strain X514 was not sequenced until recently. The metabolic pathways of the central metabolism have been studied via a ¹³ C-based metabolic pathway analysis 28. However, there is limited experimental data on the secondary metabolism, including the peripheral pathways of Thermoanaerobacter sp. strain X514. Therefore, this study chose to reconstruct a medium-scale model that included only the central metabolic pathways to demonstrate the applicability of MicrobesFlux-based FBA for understanding strain X514’s ethanol metabolism.

Using the experimental data of strain X514 reported in 28 and 29, we applied MicrobesFlux and reconstructed a metabolic model that described the carbohydrate metabolism and amino acids biosynthesis in strain X514. The drafted model consisted of 196 metabolites and 229 reactions, of which 162 reactions were intracellular reactions, 19 reactions were inflow/outflow reactions, 39 reactions were gap-filling reactions, and 9 reactions were used for biomass building block syntheses and biomass production. The intracellular reactions in the drafted model were derived from the genome-annotation of Thermoanaerobacter sp. strain X514 from the KEGG database, considering only the carbon and cofactor balance. The inflow/outflow reactions were introduced into the metabolic model only if the transporters of the specific substrates or extracellular products had been reported by published research. Moreover, two algorithms were employed to fill the gaps in the metabolic pathways. First, we implemented un-annotated pathways identified from the ¹³ C-assisted pathway analysis (e.g. the Re-type citrate synthase in TCA cycle) into the metabolic model of strain X514 28. Since the draft model was a simplified model focusing on carbohydrate metabolism and amino acids biosynthesis, the gaps might have been generated due to the lack of consideration of the metabolites exchange between the pathways that were included in the model (e.g. carbohydrate metabolism) and other pathways that were not included in the model (e.g. purine metabolism). Accordingly, we employed the principle of introducing the metabolite-exchange-reactions to fill the gaps. For example, UTP is involved in both carbohydrate metabolism and RNA synthesis. Since we did not include the RNA synthesis pathway in the MicrobesFlux model, we instead introduced a UTP exchange pathway to fill the gap. The gap-filling reactions were evaluated carefully to make sure that each one of them was necessary for feasible predictions of biomass production.

The biomass composition of Thermoanaerobacter sp. strain X514 is not yet available. In the drafted model, we used the reported biomass composition of a related species, Clostridium acetobutylicum 30. Besides the biomass composition, the growth-associated maintenance (GAM) and non-growth associated maintenance (NGAM) energies were found to play an important role in simulating the growth rate of organisms 30 31 32. In our model simulation, the NGAM was chosen as 7.6 mmol ATP/g DCW/h, as reported previously 32. To identify the GAM, we plotted the relationship between growth rate and GAM (Figure 4), and found the value of GAM that could match the experimental measurement (i.e. growth rate of strain X514 as 0.042 h^-1 reported in 29) to be 220.0 mmol ATP/g DCW. This fitted value is higher than the previously reported GAM value (150.0 mmol ATP/g DCW) in a thermopilic cellulose degrader (Clostridium thermocellum) 33. Such a high value of GAM in strain X514 could be associated with the much higher ethanol productivity in X514 than C. thermocellum during cell growth 34. This result is consistent with the positive correlation of ethanol yield and maintenance-energy in other microbial species 34. The strain 514 model and the simulated flux distributions can be found in Additional file 4 and Additional file 5. Based on this model, we also predicted the correlations between growth rate, ethanol production, and waste product outflow (Figure 5). The prediction shows a trade-off relationship between ethanol production and growth rate. The ethanol production can be increased by 25% (i.e. increased from 6.3 mmol/g DCW/h to 7.8 mmol/g DCW/h) while halving the growth rate (i.e. decreased from 0.045 h^-1 to 0.027 h^-1). By inhibiting the acetate production from 2.0 mmol/g DCW/h to 0.6 mmol/g DCW/h, the ethanol production under the optimal growth conditions could be improved by 33%. Therefore, we have shown the platform is able to make reasonable predictions for the biomass growth in response to metabolites synthesis.