ReMixHD’s goal is to revolutionize plastic recycling on an industrial scale. To aid in the necessary metabolic engineering and to estimate the platform's performance in its real-world application, we constructed a digital twin of the ReMixHD system. We created the first genome-scale metabolic model of P. fluorescens DSM50090, capable of dynamically simulating co-cultures. With an accuracy of up to 91% compared to measured data, we can predict the growth behavior, plastic degradation rate, and product yield of the platform. Comparing ReMixHD’s co-culturing concept to a singular strain, we predict a symbiotic approach to increase plastic degradation by 16%. In an industrial scale-up simulation, we estimate ReMixHD to outperform chemical recycling and incineration in CO2 emission while producing recombinant products in the process. With our digital twin, we establish a foundation to aid future teams in the metabolic engineering of P. fluorescens and further develop our efforts in establishing the species as a new model organism.
| Our Digital Twin can predict ... | Assumptions... | Resulting limitations... |
|---|---|---|
| co-culture evolution and stability over time. | all internal reactions are modeled in dynamic-equilibrium and are integrated numerically to give absolute metabolite concentrations. | the model can only predict lag- and exponential phase. It never reaches the stationary phase and stops abruptly once all nutrients are consumed. |
| plastic degradation rate, yield of any product, CO2 footprint, and more in industrial scale-up settings. | all reaction rates work at maximum physiological rate and are invariant to changes in educt and product concentrations. | poor prediction accuracy when simulating with media low substrate concentrations or at the end of exponential phase. |
| effects of any metabolic knock-outs and knock-Ins on the whole system's behavior. | the main and helper strain grow as fast as possible and do not hinder each other when satisfied with nutrients. | not possible to include regulatory gene effects on the metabolism like quorum sensing or growth repressors. |
| growth rates with up to 91% accuracy with over 70% of predictions within the 3σ interval of experimental data. | protein biosynthesis is described in one average reaction. | decreased yield accuracy when simulating recombinant proteins as products drastically different from the average protein length and composition of P. fluorescens. |
| growth media contain no cytotoxic chemicals and are at optimal temperature and pH. | it is not possible to model the effects of temperature and/or pH on the system. |
We aim to create a digital twin of ReMixHD to predict plastic degradation rates and product yield in an industrially upscaled system. Possible bottlenecks can be identified by modeling real metabolic processes of the organisms in silico. With continued collaboration with our wetlab team, we could improve the model while simultaneously optimizing wetlab experiments.
A metabolic model can be used to achieve this by calculating the reaction rates of crucial metabolism pathways of an organism dependent on the availability of metabolites in the medium . The reaction rates (fluxes) can be used to predict the total biomass and quantity of metabolic products. ReMix is an adaptable platform requiring the model to be capable of predicting product yields other than polyhydroxyalkanoates as well as growth on additional plastic sources. This can be achieved using a genome-scale metabolic model, where the genomic information of an organism is used to identify all involved metabolic pathways. We added dynamic flux-balance analysis (dFBA) and co-culture simulation (cFBA) to include temporal information of the organisms behavior needed to simulate co-cultures.
Every genome-scale metabolic model is based on a reaction list containing every metabolic reaction occurring in an organism. These lists are derived from manual curation of proteome data and additional information. Each reaction is in the following format:
Four additional reactions are needed to model the creation of more complex macromolecules like DNA, RNA, lipid membranes and proteins. As a simplification only the average length and monomer composition are considered in the reactions. For P. fluorescens the DNA and RNA forming reactions can be written as following:
The final growth of the strain is modeled as a reaction of all the macromolecules reacting together with ATP to form biomass (a new cell). To make genome-scale metabolic models more accurate, the reactions are divided into compartments like the cytosol, periplasm or the extracellular space to prevent crosstalk between reactions with identical metabolites that are physically separated in the organism.
After implementation, all reactions can be arranged in a stoichiometry matrix S, rows representing metabolic reactions, columns representing metabolites. The matrix elements are the stoichiometry factors νi of the metabolites or zero if a metabolite is not present in a reaction. Multiplying the Stoichiometry Matrix S by a vector containing all the metabolic reaction rates (fluxes) v results in a vector of all metabolites production/ consumption rates c in mol/h:
When simulating the behavior of the organism, the system of equations is solved for the flux vector v. This allows for calculation of the growth rate, nutrient utilization, and production rate of potential valuable molecules. Knock-ins and knock-outs can easily be simulated by adding additional reactions to the list or deleting existing ones.
However, solving this system of equations for v is impossible if the production rates c are not known. By limiting modeling to the exponential growth phase, this issue can be circumvented by making an internal dynamic equilibrium assumption: During optimal growth, it is assumed that each intracellular metabolite is in a dynamic equilibrium i.e. it is produced at the same rate as it is consumed. That means c = 0:
The desired flux vector v can now be obtained by solving for the null-space of this linear system of equations. However, such a system is often underdetermined as there are more reactions than metabolites in a genome-scale metabolic model. This results in infinite solutions for v. The solution space must be reduced to pick the physiologically correct flux vector.
By adding thermodynamic and experimental constraints, we can optimize the model:
Given the assumption that the organism always grows optimally, limited only by its environmental constraints, the flux vector with the highest biomass production rate in the reduced solution space represents the physiologically correct one.
Multiplying this flux vector with a given time interval results in the total amount of metabolites and biomass consumed or produced in said interval. This form of analysis is called flux balance analysis (FBA) and it is the basis of all genome-scale metabolic models.
In regular cases FBA is only suited for modeling a single organism. However, the principles of FBA can be extended to accommodate two or more organisms in the model. For community flux balance analysis (cFBA), the extracellular compartments of the FBA models representing the organisms in questions are merged by combining the reaction lists. All metabolites in the extracellular compartment have to be adjusted accordingly and potential crosstalk effects like toxic metabolites have to be added as well. The resulting digital twin model for ReMixHD operates in five compartments, one shared extracellular compartment as well as a periplasm and cytosol compartment for each strain respectively.
Even though FBA (and cFBA for co-cultures) are powerful tools in assessing the behavior of a cell culture, they cannot predict the evolution of the model over time as the fluxes are time-invariant. This results from the underlying assumption that the media composition is constant, which is necessary to solve a FBA but not representative of real world applications. To fully mitigate this problem, all reactions would have to be simulated as Michaelis-Menten kinetics. The lack of available Michaelis-Menten constants and maximum reaction rate values make this solution infeasible.
A way to circumvent the issue is dynamic flux balance analysis (dFBA). dFBA extends FBA by linking the time and concentration invariant internal reactions to a dynamically modeled extracellular medium without requiring Michaelis-Menten kinetics. Instead the flux vector is integrated over a set time interval to yield the absolute substrate concentrations and biomass. As the time integral cannot be solved analytically it is approximated using the midpoint method for numerical integration.
Modeling with dFBA, instead of FBA, results in a pseudo-dynamic model capable of simulating an initial lag- and exponential phases of bacterial growth. As the flux values are still invariant to substrate concentrations in the medium they cannot reduce with lowering nutrient availability. Therefore, the model never reaches the stationary phase and stops abruptly once all nutrients are consumed.
By combining the workflows described for FBA, cFBA and dFBA, we create a new simulation method, the community dynamic flux-balance analysis (cdFBA). It was used to establish a digital twin of the ReMixHD recycling platform capable of time-resolved simulation of the co-culture.
We evaluate the prediction accuracy of the ReMixHD digital twin and simulate the co-culture stability when growing on PET/PE minimal media. Additionally, we estimate the efficiency of an industrial scale version regarding CO2 emissions and product yield. We aid in the development of the platform in the wetlab by predicting possible growth control genes to be used in our regulatory operon.
The error of the ReMixHD digital twin was estimated by comparing the predicted growth rates of the flux balance analysis (FBA) model on M9 minimal media to measured rates (Fig 3.1.A). Differences between prediction and measurement range from 9% for acetate to 49% for ethylene glycol. 70% of predicted rates fall in the 3σ interval of the measured rates, with glucose and ethylene glycol differing further (3.2σ and 3.5σ respectively). No significant difference in reaction rates was observed for growth on succinate, glutamate, citrate and acetate (p>0.05).
Fig. 3.1. A) Comparison of simulated and measured growth rates on M9 minimal media. Maximum growth rates were obtained during the mid-exponential phase. Values are depicted as mean+sd. (n=9) Prediction errors (sd) result from the scaling factor uncertainty. B) CO2 emmision of P. fluorescens and E. coli during exponential aerobic growth. Values are depicted as mean+sd. The error bars could not be calculated due to missing data from Andersen et al. (1980). Prediction errors (sd) result from the scaling factor uncertainty.
The mixed plastic degradation capability of helper and main strain was qualitatively assessed by modeling co-culture bioremediation of PET/PE minimal media using cdFBA. Depolymerization rates were set to a maximum of 40 µmol/gDW/h for PET and 530 µmol/gDW/h for PE to achieve a degradation rate of 8 g/gDW/h when adjusted for molecular weight.
The simulation predicted the co-culture evolution as expected (Fig 3.2). Both helper and main strain grow exponentially on PE and PET, respectively. A diauxic shift is observed as the helper strain switches to sole ethylene glycol metabolism after the consumption of PE. The increased ethylene glycol uptake reduces the amount of ethylene glycol available as PET depolymerization rates remain constant. As PET is fully consumed the growth of helper and main strain stagnates.
Fig 3.2. Helper and main strain growth on PET and PE. Biomass concentration is given in gram dry-weight (gDW), substrate concentration in mmol. Simulation was started with an initial biomass of 1 gDW for helper and main strain, 200 mmol PET and 20 mmol PE.
To verify the observation made on PET/PE medium, co-culture behavior was simulated for minimal media containing only PE or PET. On PE minimal medium cdFBA predicted no growth of the main strain, due to it lacking the AlkB gene (Fig 3.3A). For PET minimal medium cdFBA showed fast growth for the main strain and reduced growth for the helper strain. (Fig 3.3B) The in silico data depict how the main strain benefits from the PET degradation more than the helper strain as it can metabolize both monomers (terephthalic acid and ethylene glycol), whereas the helper strain relies only on ethylene glycol.
Fig 3.3. A-B Helper and main strain growth on PE or PET. Biomass concentration is given in gram dry-weight (gDW), substrate concentration in mmol. Simulation was started with an initial biomass of 1 gDW for helper and main strain, 200 mmol PET or 20 mmol PE for the two minimal media respectively
The smart symbiosis approach was verified in silico by comparing the plastic degradation rates and total biomass production of the helper and main strain co-culture to a hypothetical mono-culture. The mono-culture consists of a singular P. fluorescens strain capable of performing both PET/PE depolymerization and fully metabolizing all generated monomers. Through simulation with cdFBA for the co-culture and dFBA for the mono-culture on PET/PE minimal medium the plastic degradation rates were calculated. A co-culture was able to degrade 0.57 % of PET and PE within one hour (Fig 3.4) , whereas the mono-culture only consumed 0.48 % (Fig 3.2) This results in a predicted 16% increase in mixed plastic degradation when grown on PET/PE minimal medium.
Fig 3.4. PET/PE degradation of the Monoculture. Biomass concentration is given in gram dry-weight (gDW), substrate concentration in mmol. The simulation was started with an initial total biomass of 1 gDW, 200 mmol PET or 20 mmol PE for the minimal medium respectively.
A 20000L Chemostat bioreactor with cells growing constantly at an optimum expression density of OD600 = 1.2 was chosen as a realistic future implementation of the ReMixHD system. Such a reactor allows for non-stop growth, so the dynamic-equillibrium assumption necessary for cFBA simulation is fulfilled. The CO2 emission rates, total biomass production and polyhydroxyalkanoate yields were calculated with cFBA and referenced to the PET-degradation flux to eliminate errors introduced in the degradation rate assumption. ReMixHD showed a 40% decrease in CO2 emission when compared to chemical recycling methods and a 23% reduction compared to incineration (Zero Waste Europe, 2022). Carbon emissions were conservatively calculated using stiochiometrically complete incineration. The co-culture produced 32kg dryweight of biomass per ton degraded PET amounting to a yield of 4.8 kg polyhydroxyalkanoates per ton degraded PET.
The effects of singular reaction knock-outs were assessed for the helper strain grown on PET/PE medium and full medium using dFBA. 76 deletions for PET/PE and 143 deletions for full medium reduced the growth rate below 1% of the initial objective value. The intersection of the two lists resulted in 44 potential auxotrophic markers to be used as the growth control gene in the PE/PET sensing operon characterized in the wetlab (see Table 3.1). We could verify the efficacy of common auxotrophic targets like knock-out of the dihydrofolate reductase and thymidylate synthetase in the nucleotide synthesis, as well as diaminopimelate decarboxylase in the leucine catabolism. Interestingly, other common auxotrophies like tryptophan synthesis knock-outs had no effect on the helper strain growth on PET/PE media. We presume that the terephthalic acid present in the medium can be converted to tryptophan via anthranilate, even though utilization as a carbon source is not possible in the helper strain.
Amino Acids:
Nucleic Acids:
Lipid metabolism:
Carbohydrate metabolism:
Cell wall synthesis:
Cofacor usage:
With the ReMixHD digital twin, we were able to create the first time-resolved co-culture genome-scale metabolic model for P. fluorescens DSM50090. We demonstrated that it is capable of high prediction accuracies in 70% of modeled media albeit on a small test data set. We hope that future teams can improve its accuracy and validation with further growth and metabolomics data. We successfully demonstrated the benefits of utilizing a co-culturing system over standardized mono-culture approaches and provide an important building block for the metabolic engineering of these more complex systems.
Additionally, possible auxotrophic markers that can be used for the operon mediated control of our helper strain growth were predicted. The importance of such predictions for new auxotrophic markers is highlighted by the fact that the common auxotrophic markers, such as tryptophan synthesis gene knock-outs, are incapable of controlling helper strain growth. This prediction is not described in the literature as it is unique to the interaction of PET depolymerization and the helper strain’s metabolism, highlighting the importance of this finding even more. Even though terephthalic acid cannot be utilized as a sole carbon source by the helper strain, it feeds into the tryptophan synthesis through a different pathway so that a single gene knock-out no longer results in tryptophan auxotrophy. Although different databases such as KEGG are available, it is still complicated to predict the effect of gene-knockouts without computational modeling.
The growth curves generated using the time dependent dFBA showed the expected general trend. However, the underlying estimated plastic degradation rates could not be experimentally determined. Modeling the plastic depolymerization rate itself is challenging as the model cannot predict the expression strength of the enzyme, which is directly linked to the degradation rate.
As our model was only validated with data from P. fluorescens in small culture volumes, growth behavior in large bioreactors cannot be derived from this model without further validation as upscaling will most likely not be as simple as multiplying the fluxes by a specific factor. It is necessary to perform more growth experiments in the laboratory with different bioreactor sizes to fine-tune the model and possibly predict a function for each parameter. The additional data then can be integrated into the dFBA model to optimize the parameters to best describe the metabolomics data.
Further improvements to dFBA simulations can be made by allowing metabolic fluxes to become substrate concentration dependent. Michaelis-Menten kinetics can be used to fix this problem, however, the vmax and Km values needed for a genome scale metabolic model do not exist. Mass action kinetics might present a valuable alternative to approximate the growth rate. Including substrate concentration dependent fluxes would allow for more accurate predictions of the transition from exponential to stationary phase.
Despite these limitations, our digital twin still presents itself as a powerful asset in metabolically engineering P. fluorescens and establishes the organism as a new chassis for future bioremediation projects.
Andersen, K. B., & von Meyenburg, K. (1980). Are growth rates of Escherichia coli in batch cultures limited by respiration? J Bacteriol, 144(1), 114-123. https://doi.org/10.1128/jb.144.1.114-123.1980
Our digital twin model was implemented using the MATLAB COBRA toolbox for genome-scale metabolic modeling. The toolbox was accessed using the COBRA package for python (Ebrahim et al., 2013).
The initial metabolic reaction list ‘ICW1057’ for P. fluorescens SBW25 was obtained from the only genome-scale metabolic model for P. fluorescens (Huang and Lin, 2019). ICW1057 reaction list was used as the foundation for helper and main strain, however, this model showed poor results as key reactions like glucose uptake were missing. By stimulating growth on minimal media with known P. fluorescens, further missing reactions and metabolites were identified (https://bacdive.dsmz.de/strain/12851). 57 reactions and 30 metabolites were added to the ICW1057 model to obtain the growth behavior for the ReMixHD strain P. fluorescens DSM50090 (see Table 2.1 and 2.2). The improved reaction list for the P. fluorescens DSM50090 model is available here.
| Nr. | ID | Name | Compartments | Strain | |
| 0 | cpd10001_c0 | Terephtalic_Acid_c0 | c0 | Main | |
| 1 | cpd10002_c0 | Poly_Ethylene_c0 | c0 | Helper | |
| 2 | cpd10003_c0 | Ethylene_Glycol_c0 | c0 | Helper | |
| 3 | cpd10004_c0 | DCD_c0 | c0 | Main | |
| 4 | cpd10005_c0 | Octanol_c0 | c0 | Main, Helper | |
| 5 | cpd10006_c0 | Hexanol_c0 | c0 | Main, Helper | |
| 6 | cpd10007_c0 | Decanol_c0 | c0 | Main, Helper | |
| 7 | cpd10008_c0 | Heptanol_c0 | c0 | Main, Helper | |
| 8 | cpd10009_c0 | Pentanol_c0 | c0 | Main, Helper | |
| 9 | cpd10010_c0 | Nonanol_c0 | c0 | Main, Helper | |
| 10 | cpd10011_c0 | Undecanol_c0 | c0 | Main, Helper | |
| 11 | cpd10012_c0 | Dodecanol_c0 | c0 | Main, Helper | |
| 12 | cpd10013_c0 | Dodecanoyl_CoA_c0 | c0 | Main, Helper | |
| 13 | cpd10014_c0 | Undecanoyl_CoA_c0 | c0 | Main, Helper | |
| 14 | cpd10015_c0 | Decanoyl_CoA_c0 | c0 | Main, Helper | |
| 15 | cpd10016_c0 | Nonanoyl_CoA_c0 | c0 | Main, Helper | |
| 16 | cpd10017_c0 | Octanoyl_CoA_c0 | c0 | Main, Helper | |
| 17 | cpd10018_c0 | Heptanoyl_CoA_c0 | c0 | Main, Helper | |
| 18 | cpd10019_c0 | Hexanoyl_CoA_c0 | c0 | Main, Helper | |
| 19 | cpd10020_c0 | Pentanoyl_CoA_c0 | c0 | Main, Helper | |
| 20 | cpd10021_c0 | Butanoyl_CoA_c0 | c0 | Main, Helper | |
| 21 | cpd00029_e0 | Acetate_e0 | e0 | Main, Helper | |
| 22 | cpd10003_e0 | Ethylene_Glycol_e0 | e0 | Main, Helper | |
| 23 | cpd10002_e0 | Poly_Ethylene_e0 | e0 | Helper | |
| 24 | cpd00224_e0 | L_Arabinose_e0 | e0 | Main, Helper | |
| 25 | cpd00036_e0 | Succinate_e0 | e0 | Main, Helper | |
| 26 | cpd00161_e0 | Threonine_e0 | e0 | Main, Helper | |
| 27 | cpd00363_e0 | Ethanol_e0 | e0 | Main, Helper | |
| 28 | cpd10001_e0 | Terephtalic_Acid_e0 | e0 | Main | |
| 29 | cpd10022_e0 | PET_e0 | e0 | Main |
| Nr. | ID | Name | Compartments | Strain | |
| 0 | EX_cpd00013_1_e0 | EX_NH3_1_e0 | -> e0 | Main, Helper | |
| 1 | IM_cpd00027 | IM_Glucose | c0 -> e0 | Main, Helper | |
| 2 | IM_cpd00033 | IM_Glycine | c0 -> e0 | Main, Helper | |
| 3 | IM_cpd00078 | IM_Tryptophane | c0 -> e0 | Main, Helper | |
| 4 | IM_cpd00100 | IM_Glycerol | c0 -> e0 | Main, Helper | |
| 5 | EX_cpd00029_e0 | EX_Acetate_e0 | -> e0 | Main, Helper | |
| 6 | IM_cpd00029 | IM_Acetate | c0 -> e0 | Main, Helper | |
| 7 | EX_cpd10003_e0 | EX_Ethylene_Glycol_e0 | -> e0 | Main, Helper | |
| 8 | IM_cpd10003 | IM_Ethylene_Glycol | c0 -> e0 | Main, Helper | |
| 9 | EX_cpd10002_e0 | EX_Poly_Ethylene_e0 | -> e0 | Helper | |
| 10 | IM_cpd10002 | IM_Poly_Ethylene | c0 -> e0 | Helper | |
| 11 | EX_cpd00224_e0 | EX_L_Arabinose_e0 | -> e0 | Main, Helper | |
| 12 | IM_cpd00224 | IM_L_Arabinose | c0 -> e0 | Main, Helper | |
| 13 | EX_cpd00036_e0 | EX_Succinate_e0 | -> e0 | Main, Helper | |
| 14 | IM_cpd00036 | IM_Succinate | c0 -> e0 | Main, Helper | |
| 15 | EX_cpd00161_e0 | EX_Threonine_e0 | -> e0 | Main, Helper | |
| 16 | IM_cpd00161 | IM_Threonine | c0 -> e0 | Main, Helper | |
| 17 | EX_cpd00363_e0 | EX_Ethanol_e0 | -> e0 | Main, Helper | |
| 18 | IM_cpd00363 | IM_Ethanol | c0 -> e0 | Main, Helper | |
| 19 | EX_cpd10001_e0 | EX_Terephtalic_Acid_e0 | -> e0 | Main | |
| 20 | IM_cpd10001 | IM_Terephtalic_Acid | c0 -> e0 | Main | |
| 21 | EX_PET_e0 | -> PET_e0 | -> e0 | Main | |
| 22 | rxn20047_c0 | PET_e0 -> Terephtalic acid + Ethylene_glycol | -> e0 | Main | |
| 23 | rxn20001_c0 | Terephtalic_Acid_c0 -> DCD_c0 | -> c0 | Main | |
| 24 | rxn20002_c0 | DCD_c0 -> Protocatechuate_c0 | -> c0 | Main, Helper | |
| 25 | rxn20003_c0 | 4_Carboxymuconolactone_c0 -> 3_oxoadipate_enol_lactone_c0 | -> c0 | Main, Helper | |
| 26 | rxn20004_c0 | Ethylene_Glycol_c0 -> Glycolaldehyde_c0 | -> c0 | Main, Helper | |
| 27 | rxn20005_c0 | Glycolaldehyde_c0 -> Glycolate_c0 | -> c0 | Main, Helper | |
| 28 | rxn20006_c0 | Poly_Ethylene -> relative splitting | -> c0 | Main, Helper | |
| 29 | rxn20007_c0 | Hexanol_c0 -> Hexanoate_c0 | -> c0 | Main, Helper | |
| 30 | rxn20008_c0 | Octanol_c0 -> octanoate_c0 | -> c0 | Main, Helper | |
| 31 | rxn20009_c0 | Decanol_c0 -> Decanoate_c0 | -> c0 | Main, Helper | |
| 32 | rxn20010_c0 | Hexanoate_c0 -> Hexanoyl_CoA_c0 | -> c0 | Main, Helper | |
| 33 | rxn20011_c0 | octanoate_c0 -> Octanoyl_CoA_c0 | -> c0 | Main, Helper | |
| 34 | rxn20012_c0 | Decanoate_c0 -> Decanoyl_CoA_c0 | -> c0 | Main, Helper | |
| 35 | rxn20013_c0 | Pentanol_c0 -> Pentanoyl_CoA_c0 | -> c0 | Main, Helper | |
| 36 | rxn20014_c0 | Heptanol_c0 -> Heptanoyl_CoA_c0 | -> c0 | Main, Helper | |
| 37 | rxn20015_c0 | Nonanol_c0 -> Nonanoyl_CoA_c0 | -> c0 | Main, Helper | |
| 38 | rxn20016_c0 | Undecanol_c0 -> Undecanoyl_CoA_c0 | -> c0 | Main, Helper | |
| 39 | rxn20017_c0 | Undecanol_c0 -> Undecanoyl_CoA_c0 | -> c0 | Main, Helper | |
| 40 | rxn20018_c0 | Dodecanoyl_CoA_c0 -> Decanoyl_CoA_c0 | -> c0 | Main, Helper | |
| 41 | rxn20019_c0 | Decanoyl_CoA_c0 -> Octanoyl_CoA_c0 | -> c0 | Main, Helper | |
| 42 | rxn20020_c0 | Octanoyl_CoA_c0 -> Hexanoyl_CoA_c0 | -> c0 | Main, Helper | |
| 43 | rxn20021_c0 | Hexanoyl_CoA_c0 -> Butanoyl_CoA_c0 | -> c0 | Main, Helper | |
| 44 | rxn20022_c0 | Butanoyl_CoA_c0 -> Acetyl_CoA_c0 | -> c0 | Main, Helper | |
| 45 | rxn20023_c0 | Undecanoyl_CoA_c0 -> Nonanoyl_CoA_c0 | -> c0 | Main, Helper | |
| 46 | rxn20024_c0 | Nonanoyl_CoA_c0 -> Heptanoyl_CoA_c0 | -> c0 | Main, Helper | |
| 47 | rxn20025_c0 | Heptanoyl_CoA_c0 -> Pentanoyl_CoA_c0 | -> c0 | Main, Helper | |
| 48 | rxn20026_c0 | Pentanoyl_CoA_c0 -> Propionyl_CoA_c0 | -> c0 | Main, Helper | |
| 49 | rxn20028_c0 | Ribulose -> Glyceraldehyd_3_phosphat | -> c0 | Main, Helper | |
| 50 | rxn20029_c0 | PQQH2 -> PQQ | -> c0 | Main, Helper | |
| 51 | rxn20031_c0 | Maltose -> 2Glucose | -> c0 | Main, Helper | |
| 52 | rxn20032_c0 | kynurenine -> anthranilat | -> c0 | Main, Helper | |
| 53 | rxn20033_c0 | Anthralinat _> Catechol | -> c0 | Main, Helper | |
| 54 | EX_Glucose_e0 | -> EX_Glucose_e0 | -> e0 | Main, Helper | |
| 55 | rxn20039_c0 | Butanoyl_CoA -> Succinat | -> c0 | Main, Helper | |
| 56 | rxn20045_c0 | Poly_ethylene -> Butanoyl_CoA | -> c0 | Main, Helper |
By adding the terephthalic metabolism, as introduced in the wetlab, to the preexisting model and linking it to the wild-type beta-ketoadipate pathway, a new model was created simulating our main strain. Following this pathway, terephthalic acid is introduced to the TCA cycle. The helper strain was established using the same base model and the PET and PE depolymerization pathways were added, as well as the AlkB mediated oxidation into n-alkanols. The introduced alkanols are naturally linked to the fatty acid metabolism of the helper strain.
The completed helper and main strain models were optimized for maximum biomass flux and cleaned of infeasible fluxes by implementing loopless FBA. The non-growth associated ATP-maintenance reaction was set to a minimal reaction rate of 46.3 mmol/gDW/h to maintain the physiologically necessary ATP usage to keep the cell alive (Chen et al., 2011). It was assumed that the flux observed in E. coli during aerobic growth is the same for P. fluorescens due to lacking data.
The comparison of measured and predicted growth rates on minimal media with one carbon source was chosen as a reliable and easy method to assess models performance. The carbon sources tested are metabolites from the citric acid cycle and central carbon metabolism.
The maximum growth rate of P. fluorescens was determined in 3 biological replicates as a baseline. 18 different carbon sources with a concentration of 1% (w/v) were tested on M9 minimal media to cover a broad spectrum of pathways and to eliminate potential diauxic growth effects.
Once the pre-cultures in LB medium reached an OD600 of 0.1, the cells were washed with M9 minimal salts without a carbon sources removing excess LB medium. The washed preculture was inoculated 1:10 in 200µl and plated on 96-well plates with M9 minimal medium. OD600 was measured every 10 min for 24 h using a TECAN Spark spectrometer and the maximum growth rate was determined using the Growthcurver package in R 4.1.3 (Sprouffske & Wagner, 2016).
The in silico growth rates were produced by setting all fluxes of exchange reactions with metabolites not contained in the medium to zero and was optimized for biomass. A direct comparison of the predicted growth rates rpredicted and the measured rates rmeasured was not possible, as the two values are proportional but not identical. By implementing a proportionality constant termed scaling factor k to account for this difference, the two datasets can be compared:
The equation was rearranged and k was adjusted so the rescaled mean sum of squares converged to zero. This step was repeated twice with the upper and lower values of the 1 sigma interval of the measured growth rates to estimate the 1σ interval of the scaling factor.
The carbon sources used for the model performace evaluation were chosen to avoid the overfitting of particular degradation pathways.
All following simulations were obtained by following this workflow: All exchange fluxes of carbon sources, not present in the simulated medium, were set to zero. Flux constraining and biomass optimization was performed. Using the scaling factor k, the whole model was scaled down to yield physiologically correct results. The scaled predicted fluxes were used for yield and degradation prediction.
The co-culture was simulated using the two metabolic models for the helper and the main strain. A shared medium was created including the limited amounts of nutrients to be observed. By subtracting the consumed nutrient values of both strains from the initial medium values, biomass growth was linked to depletion of nutrients. The growth of the biomass is terminated when all energy sources in the medium are consumed. Predictions with cFBA followed the same workflow described for FBA, however, the objective function was altered to optimize for the sum of the biomass fluxes of helper and main strain.
For dFBA simulations, an initial medium composition of 200 mmol carbon source (20 mmol for polymer degradation) was set. Other necessary non-carbon metabolites (e.g oxygen, carbondioxide, water, etc.) were set to 10 mol to ensure that the growth simulation is not limited by them. The fluxes are calculated and multiplied by a small, constant time interval to give the approximate change in media composition over the time frame. A time step of 0.1 h was found to give good resolution with moderate calculation time. Subtracting this change from the initial medium composition gives a new initial medium composition. The process is repeated until the substrate concentration in the medium reaches negative values. At each timestep, a vector of media concentration is stored for plotting and analysis.
dFBA of the coculture model was performed similarly to the dFBA of a singular model. However, the optimization step at each time point included two optimizations on the same medium composition. One optimization calculates the biomass and the substrate effects for the main strain, the other for the helper strain. The media compositions were updated accordingly and stored for later analysis.
To identify the metabolic rations with the highest impact on the helper strains growth, its biomass flux was simulated on PET/PE minimal medium and in an unlimited medium containing all possible extracellular metabolites. Then each reaction was knocked-out individually and the effect on the PET/PE and full medium growth recorded. Knockouts that resulted in less than 1% of initial growth rate on both media were deemed essential and present good candidates for controlling the helper strain’s growth.
A fuction to scale-up processes was developed to assess the future potential application of the ReMixHD system. This method is capable of simulating a chemostat (also known as a continuous bioreactor), which maintains constant media composition and cell density, allowing for consistent cell growth at maximum growth rate.
Our scale-up model utilizes the bioreactor's volume, elapsed time, and substrate properties to predict the degradation of various substrates, the production of biomass, emitted CO2, and polyhydroxyalkanoates (PHA). This extrapolation function is based on the following assumptions: A cell of P. fluorescens has a comparable dry weight to E. coli, estimated at 3 x 10-13 grams (Neidhardt et al., 1996). In the bioreactor, a dominant cell density of OD = 1.2 is observed. Assuming that 0.1 OD units correspond to 7 x 105 cells/ml, this results in a cell count of 8.4 x 109 cells/l.
Next, the total weight of dry biomass in the bioreactor can be calculated. From there, the determination of the biomass produced during a degradation process depends on the growth rate and duration. The generated PHA mass is derived from the potential storage capacity of PHA in P. fluorescens, which is reported to be 15% of the cell dry weight according to (Vladu et al., 2019).
Subsequently, the quantity of the degraded substrate was calculated. This was achieved by multiplying the substrate's uptake rate with CDWbioreactor, t, and the molecular weight of our substrate.
To compute the emitted CO2, the model uses the molecular weight of CO2 as a multiplier for its flux. This value is multiplied by CDWbioreactor and t.
This model serves as an extrapolation for industrial-scale predictions of degradation rates, product yields, and CO2 emissions. It is extensible to include the status of co-cultures, such as the proportion of helper and main strain.
Chen, X., Alonso, A. P., Allen, D. K., Reed, J. L., & Shachar-Hill, Y. (2011). Synergy between 13C-metabolic flux analysis and flux balance analysis for understanding metabolic adaption to anaerobiosis in E. coli. Metabolic Engineering, 13(1), 38–48. https://doi.org/10.1016/j.ymben.2010.11.004
Ebrahim, A., Lerman, J. A., Palsson, B. O., & Hyduke, D. R. (2013). COBRApy: COnstraints-Based Reconstruction and Analysis for Python. BMC systems biology, 7, 74. https://doi.org/10.1186/1752-0509-7-74
Huang, X. and Lin, Y.-H. (2020). Reconstruction and analysis of a three-compartment genome-scale metabolic model for Pseudomonas fluorescens. Biotechnology and Applied Biochemistry, 67, 133-139. https://doi.org/10.1002/bab.1852
Sprouffske, K., Wagner, A. (2016). Growthcurver: an R package for obtaining interpretable metrics from microbial growth curves. BMC Bioinformatics 17, 172. https://doi.org/10.1186/s12859-016-1016-7
Neidhardt et al. (1996). Escherichia coli and Salmonella typhimurium, 2nd edition..
Vladu, M.-G., Petrescu, M., Stefaniu, A., Savoiu, G., Spiridon, M., Eremia, M., & Irina, L. (2019). Studies on polyhydroxyalkanoates biosynthesis by some Pseudomonas spp. strains. Romanian Biotechnological Letters, 24, 388-394. https://doi.org/10.25083/rbl/24.3/388.394