Team:Edinburgh UG/Model

Model

Summary

We predicted H2 production rate in four E. coli strains (wild type, SHI, HydA and FDH) on three different carbon sources (glucose, galactose, lactate) using constraint-based modelling. In most cases, FDH strain showed the highest H2 production rate in our simulation results, which influenced our team to express it in E. coli experiments. SHI strain was not taken into consideration in experiments in the end, because its hydrogen production rate was lower than wild type or at the same level. Lactate (representing pot ale in our model) was the carbon source with the smallest H2 production rate, so we decided not to use it in our latest E. coli experiments. Simulation attempts on Rhodobacter sphaeroides were also performed, but their results remain inconclusive.

Introduction to Constraint-based Modelling

We first learnt about constraint-based modelling from the wiki of Edinburgh University Overgraduate 2018 team [1]. We chose this simulation technique for our project, because it is widely applied in biotechnology and metabolic engineering for strain design to improve a product yield [2]. It has also been successful with its predictions matching the experimental data [3]. In addition, bacteria that we considered working on, E. coli and Rhodobacter sphaeroides, had a wealth of information about their metabolisms, which is crucial in this type of modelling [4, 5]. Besides that, it is straightforward to add and delete enzymatic reactions which would correspond to our genetic modifications.

One of the most popular and primary methods in constraint-based modelling is Flux Balance Analysis (FBA). It relies on a genome-scale metabolic network reconstruction of a cell, which contains all information on what metabolic reactions take place within a given cell and what enzymes facilitate them. It also includes what genes are associated with these reactions. However, such a reconstruction usually does not include transcriptome or proteome details. FBA does not require any kinetic data, which is an advantage, because such large whole-cell systems would have to deal with way too many parameters [6]. It is based on stoichiometric values of reactions and their fluxes instead. From the mathematical point of view it can be represented by this equation:

S is an m by n matrix, where m denotes the number of compounds in a cell and n is the number of all reactions. In this case each row of the matrix represents one compound and a column is for one reaction. So a matrix element contains a stoichiometric value indicating if a given compound takes part in a given reaction. A positive number shows that a compound is a product. Conversely, a negative number means that a compound is a substrate. Zero would mean that a specific metabolite does not take part in a given reaction. Going further, vector v is composed of n reaction fluxes. A flux means a reaction rate in this context and it is expressed in mmol gDW-1h-1 (gDW – gram dry weight) [2].

The main assumptions of FBA are that the whole biological system is at steady state, Sv = 0. This implies that there is no internal accumulation or consumption of the biochemical substances [7]. Experimentally, this assumption holds in the exponential phase of bacterial growth only in batch cultures. Alternatively, chemostats can be used to maintain constant growth rate and make continuous cultures [8]. Using this steady-state constraint, we could find one solution that optimises an objective of the cellular system with help of linear programming. The objective can be understood as the cell’s main goal – in our case it could be the maximization of biomass growth rate or biohydrogen production rate which we aim to improve. Another more general assumption we will make here is that the cell wants to optimize its growth rate, which is chosen as an objective quite often in FBA literature [9]. The results of the FBA optimization are the maximum theoretical value of the objective and possible flux values for all reactions which satisfy the given objective [2]. Since this model focuses mainly on fluxes and represents only one point in time, it is not possible to explicitly derive metabolic concentrations from the results [10].

Another thing that needs to be noted here is the fact that there is a whole range of possible flux values for one reaction, but FBA shows only one value. However, there is a unique value for the objective [2]. To minimize the flux solution space, more bounds (constraints) are applied on reaction fluxes:

They indicate smallest and largest possible fluxes for a given reaction. The bounds also represent the directionality of a reaction. If li is negative and ui is positive, then we have a reversible reaction. If any of these are zero, then the reaction goes only in one direction. Similarly, ui = li = 0 means the reaction does not take place. The most straightforward way to modify constraints in this model is by changing bounds of exchange reactions. This is a special type of reactions that symbolises introduction or removal of substances into the system, which is composed of the cell and its (closest) surroundings. That is why we decided to determine the best growth media for our genetically modified bacteria. These bounds are also the easiest to measure experimentally, so we managed to find uptake rates of the main carbon sources for E. coli and R. sphaeroides in the literature. An uptake rate for an exchange reaction is represented by its lower bound with a negative value by convention [2].

We focused only on comparing minimal growth media with a single carbon source, because FBA does not predict well when more sugars come into play. It inflates the results and does not take the gene regulation and carbon source preferences of E. coli into consideration. The bacterium prefers to consume glucose first and then another sugar, however FBA results from scientific literature show that both sugars are taken up simultaneously. Such a discrepancy in data occurs, because these microbial metabolic reconstructions do not contain information about regulatory networks. However, researchers slowly started developing models that include this [11], but they are not considered in this project.

Escherichia coli Simulations

Modelling Design

For E. coli simulations, reconstruction iJO1366 was used. The modelling was carried out using COBRApy Python package. The E. coli reconstruction is included in this library and it has bounds for the most relevant mineral uptake already preset, which represent part of the minimal growth medium. NH4 is the only nitrogen source here. E. coli produces hydrogen in anaerobic conditions [7], so oxygen uptake was removed by setting its lower and upper bounds to zero. There were more bounds for internal reactions altered to portray anaerobic conditions better. They were based on the paper written by J. Seppälä and her colleagues [7]. The reaction catalysed by formate hydrogen lyase was made fully reversible. Reactions facilitated by pyruvate and 2-oxoglutarate dehydrogenases were deactivated. Additionally, the action of isocitrate dehydrogenase was made irreversible to favour isocitrate production.

The uptake rate in E. coli for glucose in anaerobic conditions is 18.5 mmol gDW-1h-1 [12]. It was approximated and set to 20 in the simulations for the sake of simplicity. Uptake rates for other carbon sources were also set to -20 mmol gDW-1 h-1, which is an arbitrary value based on typical substrate uptake rates and quite common in FBA studies [13, 4]. We selected glucose for this work, because it is a common sugar used as a growth medium for E. coli. Similar reasoning was in the case of galactose.

Additionally, we decided to integrate human practices into our research and try growing our bacteria on whisky by-products (in particular pot ale which is a liquid by-product of distillation reaction rich in lactate, propionate and acetate [14]). That is why we considered this medium in modelling. We ended up representing pot ale only as lactate, because E. coli simulations in anaerobic conditions did not work at all with propionate and acetate as single carbon sources – there was no growth.

In the simulations, we considered 3 genetically modified strains of E. coli apart from the wild type. The reactions we added to the WT E. coli reconstruction for each strain are specified below:

1st strain

FNOR (ferredoxin-NAD+ reductase from Trichomonas vaginalis)

NADH + 2 reduced ferredoxin ⇔ NAD+ + 2 oxidised ferredoxin + H+

FNR (ferredoxin-NADP+ reductase from Chlamydomonas reinhardtii)

NADPH + 2 reduced ferredoxin ⇔ NADP+ + 2 oxidised ferredoxin + H+

HydA ([FeFe]-hydrogenase from Chlamydomonas reinhardtii)

2 reduced ferredoxin + 2 H+ ⇔ H2 + 2 oxidised ferredoxin

2nd strain

SHI ([NiFe]-hydrogenase from Pyroccocus furiosus)

NADPH + 2 H+ ⇔ NADP+ + H+

3rd strain

FDH (formate dehydrogenase from Clostridium carboxidivorans)

CO2 + NADH ⇔ formate + NAD+

The first strain with HydA hydrogenase was an inspiration taken from the Macquarie 2017 iGEM team, which implemented this enzyme for their project successfully [15]. The Macquarie team expressed FNR and HydA enzymes in E. coli. FNOR enzyme was our own suggestion. These enzymes utilise ferredoxin. This protein has not been found in the iJO1366 reconstruction, even though it has been proved experimentally to appear in E. coli naturally [16]. For this reason, we have included ferredoxin in our reconstruction. More details on why these enzymes were considered by our team can be found in Design section.

Results

Figure 1 and figure 2 show our first FBA results for biohydrogen production. These simulations were carried out for two different objectives – biomass growth and H2 exchange reaction. Hydrogen exchange flux was set as an objective to show the highest production rate possible and test maximum theoretical capabilities of E. coli metabolism. Biomass growth rate was also considered as an objective, because we made an assumption earlier that E. coli wants to optimise its growth.

As it can be noticed in figure 1, FDH outperforms all other strains. However, in lactate the difference between FDH and HydA strains is much smaller. In contrast, SHI strain has much lower hydrogen production rate than wild type when grown on glucose and galactose. It also appears that galactose is a minimally better carbon source than glucose. Lactate is the poorest medium in all cases.

In figure 2, hydrogen flux is at least 3 times higher than in figure 1 for some instances, which is not surprising given the fact that H2 exchange is set as an objective here. In this situation, biomass growth rate is 0, so experimentally it is unlikely for the bacterium to allocate all of its resources to hydrogen production and stop growing. FDH is still better than wild type, but at this point it can produce as much hydrogen as HydA strain.


Figure 1. H2 flux for various E. coli strains growing on different media with a single carbon source using FBA. Biomass growth rate was set as an objective.

Figure 2. H2 flux for various E. coli strains growing on different media with a single carbon source using FBA. H2 exchange was set as an objective.

When switching between carbon sources for FBA simulations, we noticed that we were getting varying hydrogen flux values for the same growth medium. To avoid this situation, it is advisable to use model.copy() command (when coding in Python). However, this occurrence made us use other techniques such as flux variability analysis. As it was mentioned earlier, there are many ways to obtain the same objective value in FBA and from this technique one can see only one solution. To learn the range of all possible values, flux variability analysis (FVA) is applied instead. FVA is technically an FBA extension [17]. Its results can be found in figures 3 and 4.

In figure 3, we can see that WT and SHI strains exhibit a range of feasible flux values, meanwhile FDH and HydA strains center at only one point. It implies that WT and SHI strains can possibly have as good production rate as HydA and get close to FDH strain production rate. Figure 4 basically shows the same results as figure 2. This plot reiterates the fact that an objective has only one unique value after FBA optimization.


Figure 3. Minimum and maximum possible H2 fluxes for various E. coli strains growing on different media with a single carbon source using FVA. Biomass growth was set as an objective.

Figure 4. H2 flux for various E. coli strains growing on different media with a single carbon source using FBA. H2 exchange was set as an objective. These FVA results look exactly like the FBA output from figure 2.

Attempts in Rhodobacter sphaeroides Modelling

Since Rhodobacter sphaeroides was a central and prevalent part of our project experimentally for quite a long time, we attempted predicting hydrogen production using constraint-based modeling in this bacterium as well. The iRsp1140 reconstruction was used to perform FBA. The simulations were carried out using Cobra Toolbox in Matlab, because this reconstruction (or any other version referring to this bacterium) did not work in COBRApy – always giving zero fluxes in the whole system, even though all relevant reaction bounds were open.

Our team considered developing two new strains to improve H2 production rate – one would introduce HydA with ferredoxin reductases and the other one would express SHI gene. The added reactions to the WT reconstruction are exactly the same as in the case of E. coli (strains 1 and 2). This time ferredoxin was already included in iRsp1140 [5]. It is proved experimentally that the best carbon and nitrogen sources for highest H2 production in R. sphaeroides are succinate and glutamate respectively [18], so we started conducting simulations in minimal Siström’s medium with these substances. We selected constituents for Siström’s medium based on the article written by B. Burger and his colleagues [19]. The uptake rates for succinate and glutamate were set to 3 and 1 mmol gDW-1h-1 respectively in agreement with experimental values as it was done in the paper authored by S. Imam and other researchers [20].

It has been observed that that all strains (SHI, HydA and WT) had H2 flux of 12.5162 mmol gDW-1h-1 with H2 exchange as an objective. This lack of difference in hydrogen production rate could be explained by the fact that the new reactions we added already exist in R. sphaeroides reconstruction. It made us realise that in basic FBA duplicating a reaction in the model does not account for the increased concentration of the catalysing enzyme. In this case, proteomic constraints with enzymatic concentrations could be incorporated in FBA to see if it actually helps in introducing a hydrogen flux difference between these strains [21]. Unfortunately, it is quite a lengthy process to collect all the enzymatic data for this model, so it has not been fully carried out and remains as a for potential future research on this project.

How Modelling Influenced the Experimental Design

At the beginning of our project, simulation results showed that incorporation of SHI hydrogenase in E. coli was not worth pursuing, so it made our team revert focus to other solutions such as Rhodobacter sphaeroides. Unfortunately, simulations on this bacterium did not provide reliable results due to technical issues in COBRAPy and the need for more enzymatic constraints. However at the later stages of our summer research we encountered FDH enzyme in scientific literature [22], so we decided to predict its contribution to H2 production rate in silico. These data assured us using FDH in E. coli can be a good time investment for our research, because we found out it has the highest hydrogen production rate among all the tested strains. Also, the results made us realise pot ale (represented as lactate in the simulations) is a poorer growth source than other sugars, so we decided not to use whisky by-products as nutrient medium for E. coli at the latest stage of our project to save us time and help achieve our full potential in the improvement of H2 production.

Potential Improvements

The simulations for this project can be further developed in many ways. For example, the choice of objective for FBA in E. coli could be reconsidered. Even though biomass growth rate is commonly selected by researchers as an objective [9], it is not the best option for these bacteria in every biological setting [23]. Additionally, uptake rates could be measured experimentally for carbon sources such as lactate or galactose to put more realistic bounds on the reconstruction. The E. coli model would also benefit a lot from proteomic constraints with enzymatic concentrations (which was already mentioned in the context of R. sphaeroides simulations) as well as FBA models with regulatory information [11]. As far as it goes for R. sphaeroides, it would be useful to explore further its hydrogen production with biomass growth as an objective using FVA and FBA. Incorporation of regulatory and proteomic constraints would also help provide more accurate results.

The scripts used for the simulations can be found here.

References

  1. Overgraduate Edinburgh University iGEM 2018 team. https://2018.igem.org/Team:Edinburgh_OG/Model. Accessed June 4, 2019.
  2. Orth JD, Thiele I, Palsson BO. What is flux balance analysis?. Nature biotechnology.2010;28(3):245–248. doi:10.1038/nbt.1614.
  3. Edwards JS, Ibarra RU, Palsson BO. In silico predictions of Escherichia coli metabolic capabilities are consistent with experimental data. Nature Biotechnology. 2001;19(2):125-30. doi:10.1038/84379.
  4. Orth JD, Conrad TM, Na J, et al. A comprehensive genome-scale reconstruction of Escherichia coli metabolism--2011. Mol Syst Biol. 2011;7:535. Published 2011 Oct 11. doi:10.1038/msb.2011.65.
  5. Imam S, Noguera DR, Donohue TJ. Global insights into energetic and metabolic networks in Rhodobacter sphaeroides. BMC Systems Biology. 2013;7(89).
  6. Seaver SMD, Sales-Pardo M, Guimerà R, Amaral LAN. Phenomenological Model for Predicting the Catabolic Potential of an Arbitrary Nutrient. PLOS Computational Biology. 2012;8(11): e1002762. doi:10.1371/journal.pcbi.1002762.
  7. Seppälä JJ, Larjo A, Aho T, Yli-Harja O, Karp MT, Santala V. Prospecting hydrogen production of Escherichia coli by metabolic network modeling. International Journal of Hydrogen Energy. 2013;28(27): 11780-11789. doi:10.1016/j.ijhydene.2013.07.002.
  8. Stanbury PF, Whitaker A, Hall SJ. Principles of fermentation technology. Butterworth-Heinemann; 2016:31-35.
  9. García Sánchez CE, Torres Sáez, RG. Comparison and analysis of objective functions in flux balance analysis. Biotechnol Progress, 2014;30:985-991. doi:10.1002/btpr.1949.
  10. Bordbar A, Monk JM, King ZA, Palsson BO. Constraint-based models predict metabolic and associated cellular functions. Nat Rev Genet. 2014 Feb;15(2):107-20. doi:10.1038/nrg3643. Epub 2014 Jan 16.
  11. King ZA, Lloyd CJ, Feist AM, Palsson BO. Next-generation genome-scale models for metabolic engineering. Curr Opin Biotechnol. 2015 Dec;35:23-9. doi:10.1016/j.copbio.2014.12.016. Published 2015 Jan 7.
  12. Varma A, Palsson BO. Stoichiometric Flux Balance Models Quantitatively Predict Growth and Metabolic By-Product Secretion in Wild-Type Escherichia coli W3110. Appl Environ Microbiol. 1994;60(10):3724–3731.
  13. Sridhara V, Meyer AG, Rai P, et al. Predicting growth conditions from internal metabolic fluxes in an in-silico model of E. coli. PLoS One. 2014;9(12):e114608. Published 2014 Dec 12. doi:10.1371/journal.pone.0114608.
  14. Graham J, Peter B, Walker GM, Wardlaw A, Campbell E. Characterisation of the pot ale profile from a malt whisky distillery. Distilled spirits: science and sustainability: proceedings of the Worldwide distilled spirits conference. 2012;4.
  15. Macquarie University iGEM 2017 team. https://2017.igem.org/Team:Macquarie_Australia. Accessed June 25, 2019.
  16. UniProtKB - P0A9R4 (FER_ECOLI). Uniprot. https://www.uniprot.org/uniprot/P0A9R4#. Accessed July 9, 2019.
  17. Vlasov V. Flux Variability Analysis. The COBRA Toolbox. https://opencobra.github.io/cobratoolbox/stable/tutorials/tutorialFVA.html. Accessed August 10, 2019.
  18. Chongcharoenthaweesuk P. Hydrogen Production by Rhodobacter sphaeroides and its Analysis by Metabolic Flux Balancing [doctoral dissertation]. The University of Manchester. 2013.
  19. Burger BT, Saheed I, Scarborough MJ, Noguera DR, Donohue TJ. Combining Genome-Scale Experimental and Computational Methods To Identify Essential Genes in Rhodobacter sphaeroides. mSystems. 2017;2(3). doi:10.1128/mSystems.00015-17.
  20. Imam S, Yilmaz S, Sohmen U, et al. iRsp1095: a genome-scale reconstruction of the Rhodobacter sphaeroides metabolic network. BMC Syst Biol. 2011;5:116. doi:10.1186/1752-0509-5-116.
  21. Sánchez BJ, Zhang C, Nilsson A, Lahtvee PJ, Kerkhoven EJ, Nielsen J. Improving the phenotype predictions of a yeast genome-scale metabolic model by incorporating enzymatic constraints. Mol Syst Biol. 2017;13(8):935. doi:10.15252/msb.20167411.
  22. Alissandratos A, Kim HK, Matthews H, Hennessy JE, Philbrook A, Easton CJ. Clostridium carboxidivorans strain P7T recombinant formate dehydrogenase catalyzes reduction of CO(2) to formate. Appl Environ Microbiol. 2013;79(2):741–744. doi:10.1128/AEM.02886-12.
  23. Schuetz R, Kuepfer L, Sauer U. Systematic evaluation of objective functions for predicting intracellular fluxes in Escherichia coli. Mol Syst Biol. 2007;3:119. doi:10.1038/msb4100162.

Follow us on:
Facebook: HydrolyteiGEM
Facebook: @Edniburgh_iGEM
Facebook: @edniburgh_igem