(659a) An Integrated Approach to Maximize Process and Strain Efficiency in Biochemical Processes:the Case of Bioplastics Production and Muconic Acid.
AIChE Annual Meeting
2022
2022 Annual Meeting
Computing and Systems Technology Division
Sustainable Process Synthesis, Design and Intensification
Thursday, November 17, 2022 - 3:30pm to 3:51pm
Empowered by modern biology, industrial biotechnology is generating unprecedented amounts of tools and
knowledge that can be funneled into bioprocess development pipelines. Development stages include the generation of
de novo biosynthetic pathways experimental work to build enzymes, experimental work to test enzymes, and the
design of process engineering units to produce chemicals at pilot and industrial scales. An integrated approach to
simultaneously address strain and process design is not only able to coordinate the selection of degrees of freedom for
maximum efficiency but also bridge gaps between theoretical and experimental work. Computational strain design is
made however in the absence of real-time experiments and real-time dynamics that make scale-up studies uncertain
and prone to failure. Significant inconsistencies between model growth prediction with experiments reflect the most
critical gaps between synthetic biology theory and engineering, ones essentially between theoretical (curated) and
process engineering (observed) dynamics. Such an integrated approach is particularly important in the production of
plastics as they are currently based, almost entirely, on fossil-based feedstocks. Each year an average of 78 million of
non-renewable feedstocks are consumed to produce plastics that end up - in approximately equal shares - to oceans,
incinerators, and landfills. The use of renewable feedstocks makes an important alternative to replace fossil feedstocks
but is challenged to match the high efficiency of conventional polymer processes.
The paper contributes with a systems methodology designed to integrate design models made up by curated
(theoretical) and observed (real-life) kinetics. The working hypothesis is that given is a set of (under-defined) genomescale
models (GSMs) curable to produce extensive sets of strain kinetics. Instead, process engineering models involves
a much more restricted set of observations supported by experimental data (Fig. 1). The systems approach is intended
to systematically connect the two domains (theoretical, experimental) with a methodology that can be applied for the
general case. The proposed methodology involves a sequence of clearly defined steps. The system stoichiometry is
first defined through biochemical and experimental data; they are both integrated to the strain model as constraints.
Steady state fluxes and calculated based on a combined use of thermodynamics and metabolomics. Once conserved
moieties are identified, metabolites are classified between dependent and independent ones. Kinetic parameters are
calculated for every reaction so that to fit steady state fluxes based on mechanistic kinetics expressions. Consistency
checks and pruning are used to address system stability and consistency with experiments. Flux control coefficients
for desired metabolite fluxes are calculated using metabolic control analysis (MCA). The work uses clustering and
advanced statistical analysis on the control coefficient population is performed to determine the impact of key enzymes
on the desired flux.
The approach is demonstrated with the production of muconic acid, a high value product that gathers interest in the
production of new resins, bio plastics, food additives, agrochemicals, and pharmaceuticals. The stoichiometry of the
system has used a yeast genome scale model. Three heterologous reactions (PaAroZ, KpAroY, CaCatA) were added
to the GSM via shikimate pathway branching (Fig. 2). Τhe model consisted of 306 reactions and 300 metabolites. It
was curated and next reduced using data reduction (redGEM) and classification algorithms (lumpGEM) at EPFL.
Experimental data are integrated into the reduced model as well as a set of appropriate heterologous reactions. The
solution space from Thermodynamic Flux Balance Analysis was sampled as well as metabolite concentrations and
reaction fluxes profiles so that to agree with thermodynamics. Kinetic parameters for all the reactions are calculated
using Monte Carlo simulations; for each sampled set kinetic parameters are calculated to verify the particular state.
The produced kinetic models underwent pruning and stability checks. Examples of model populations are shown in
Fig. 3. A total of 23,500 potential kinetic models were generated. Out of them 372 models (1.58%) agreed with the
experimental data and passed successfully the pruning stage. Pruning reduces the models to those that are
physiologically relevant and consistent with the experimental data and stability checks identifies the models that show
great stability to a wide range of random perturbations. Stable models are used to calculate flux control coefficients
and concentration control coefficients. The work continued with the stability test which consisted of 2 sets of 100
random perturbations each on two different ranges, a shorter and a wider one. As expected, models performed better
in the shorter range as they had higher stability scores. Moreover, some kinetic models reached new steady states; by
imposing a clustering algorithm the work calculated the new centroids of the new steady states. An interesting
observation is that, in their majority, kinetic models that are based on the same TFA sample reached the same
alternative steady state.
Once kinetic models that demonstrate robust cellular behavior are selected, a decision tree analysis (Fig. 4) was
performed on the kinetic parameters with the a purpose to determine parameters that affect stability highly. For a
stability score of 90% the most important kinetic parameter revealed to include the thermodynamic displacement for
SUCCt (Gamma<0.389) and the nad_c saturation for reaction ALCD26xi (sigma>0.725). To increase the percentage
of stable models during the generation step, one could employ those bounds on those two parameters. Iterative cycles
would lead to more stable kinetic models and eventually constrain a plethora of additional parameters. Overall, 29
models out of the 372 (0.12%) managed to pass consistency checks and demonstrate stability in random perturbations.
The 29 models indicate key enzymes affecting muconic acid flux and process efficiency; they also point to the
potential scope for improvement and increase in process efficiency. For enzymes that had large control coefficients
for muconic fluxes we performed perturbations to calculate their impact. The candidate enzymes (PGI, TKT2, PGL,
DHQTi) showed an increase in muconic flux with varying results. PGI downregulation by 0.5 resulted in 12.5 times
bigger than reference muconic flux whereas TKT2 PGL downregulation by 0.5 resulted in 1.4 and 1.3 times
accordingly bigger flux. DHQTi upregulation by 1.5 resulted in 1.8 times bigger than reference muconic flux. We also
tested upregulating the heterologous enzymes (PaAroz,KpAroy) of the muconic pathway but the increase in muconic
flux was insignificant.
Although demonstrated on muconic acid, the approach follows generic steps of analysis that could benefit from
advanced analytics, data engineering and AI methods. It could also be connected with superstructure optimization
letting degrees of freedom quantify benefits for the integrated flowsheet, not by means of reaction yields and
selectivity. Work in progress includes the application of the methodology in the development of in-situ product
recovery systems (e.g. development of innovative reactive-separation systems).
knowledge that can be funneled into bioprocess development pipelines. Development stages include the generation of
de novo biosynthetic pathways experimental work to build enzymes, experimental work to test enzymes, and the
design of process engineering units to produce chemicals at pilot and industrial scales. An integrated approach to
simultaneously address strain and process design is not only able to coordinate the selection of degrees of freedom for
maximum efficiency but also bridge gaps between theoretical and experimental work. Computational strain design is
made however in the absence of real-time experiments and real-time dynamics that make scale-up studies uncertain
and prone to failure. Significant inconsistencies between model growth prediction with experiments reflect the most
critical gaps between synthetic biology theory and engineering, ones essentially between theoretical (curated) and
process engineering (observed) dynamics. Such an integrated approach is particularly important in the production of
plastics as they are currently based, almost entirely, on fossil-based feedstocks. Each year an average of 78 million of
non-renewable feedstocks are consumed to produce plastics that end up - in approximately equal shares - to oceans,
incinerators, and landfills. The use of renewable feedstocks makes an important alternative to replace fossil feedstocks
but is challenged to match the high efficiency of conventional polymer processes.
The paper contributes with a systems methodology designed to integrate design models made up by curated
(theoretical) and observed (real-life) kinetics. The working hypothesis is that given is a set of (under-defined) genomescale
models (GSMs) curable to produce extensive sets of strain kinetics. Instead, process engineering models involves
a much more restricted set of observations supported by experimental data (Fig. 1). The systems approach is intended
to systematically connect the two domains (theoretical, experimental) with a methodology that can be applied for the
general case. The proposed methodology involves a sequence of clearly defined steps. The system stoichiometry is
first defined through biochemical and experimental data; they are both integrated to the strain model as constraints.
Steady state fluxes and calculated based on a combined use of thermodynamics and metabolomics. Once conserved
moieties are identified, metabolites are classified between dependent and independent ones. Kinetic parameters are
calculated for every reaction so that to fit steady state fluxes based on mechanistic kinetics expressions. Consistency
checks and pruning are used to address system stability and consistency with experiments. Flux control coefficients
for desired metabolite fluxes are calculated using metabolic control analysis (MCA). The work uses clustering and
advanced statistical analysis on the control coefficient population is performed to determine the impact of key enzymes
on the desired flux.
The approach is demonstrated with the production of muconic acid, a high value product that gathers interest in the
production of new resins, bio plastics, food additives, agrochemicals, and pharmaceuticals. The stoichiometry of the
system has used a yeast genome scale model. Three heterologous reactions (PaAroZ, KpAroY, CaCatA) were added
to the GSM via shikimate pathway branching (Fig. 2). Τhe model consisted of 306 reactions and 300 metabolites. It
was curated and next reduced using data reduction (redGEM) and classification algorithms (lumpGEM) at EPFL.
Experimental data are integrated into the reduced model as well as a set of appropriate heterologous reactions. The
solution space from Thermodynamic Flux Balance Analysis was sampled as well as metabolite concentrations and
reaction fluxes profiles so that to agree with thermodynamics. Kinetic parameters for all the reactions are calculated
using Monte Carlo simulations; for each sampled set kinetic parameters are calculated to verify the particular state.
The produced kinetic models underwent pruning and stability checks. Examples of model populations are shown in
Fig. 3. A total of 23,500 potential kinetic models were generated. Out of them 372 models (1.58%) agreed with the
experimental data and passed successfully the pruning stage. Pruning reduces the models to those that are
physiologically relevant and consistent with the experimental data and stability checks identifies the models that show
great stability to a wide range of random perturbations. Stable models are used to calculate flux control coefficients
and concentration control coefficients. The work continued with the stability test which consisted of 2 sets of 100
random perturbations each on two different ranges, a shorter and a wider one. As expected, models performed better
in the shorter range as they had higher stability scores. Moreover, some kinetic models reached new steady states; by
imposing a clustering algorithm the work calculated the new centroids of the new steady states. An interesting
observation is that, in their majority, kinetic models that are based on the same TFA sample reached the same
alternative steady state.
Once kinetic models that demonstrate robust cellular behavior are selected, a decision tree analysis (Fig. 4) was
performed on the kinetic parameters with the a purpose to determine parameters that affect stability highly. For a
stability score of 90% the most important kinetic parameter revealed to include the thermodynamic displacement for
SUCCt (Gamma<0.389) and the nad_c saturation for reaction ALCD26xi (sigma>0.725). To increase the percentage
of stable models during the generation step, one could employ those bounds on those two parameters. Iterative cycles
would lead to more stable kinetic models and eventually constrain a plethora of additional parameters. Overall, 29
models out of the 372 (0.12%) managed to pass consistency checks and demonstrate stability in random perturbations.
The 29 models indicate key enzymes affecting muconic acid flux and process efficiency; they also point to the
potential scope for improvement and increase in process efficiency. For enzymes that had large control coefficients
for muconic fluxes we performed perturbations to calculate their impact. The candidate enzymes (PGI, TKT2, PGL,
DHQTi) showed an increase in muconic flux with varying results. PGI downregulation by 0.5 resulted in 12.5 times
bigger than reference muconic flux whereas TKT2 PGL downregulation by 0.5 resulted in 1.4 and 1.3 times
accordingly bigger flux. DHQTi upregulation by 1.5 resulted in 1.8 times bigger than reference muconic flux. We also
tested upregulating the heterologous enzymes (PaAroz,KpAroy) of the muconic pathway but the increase in muconic
flux was insignificant.
Although demonstrated on muconic acid, the approach follows generic steps of analysis that could benefit from
advanced analytics, data engineering and AI methods. It could also be connected with superstructure optimization
letting degrees of freedom quantify benefits for the integrated flowsheet, not by means of reaction yields and
selectivity. Work in progress includes the application of the methodology in the development of in-situ product
recovery systems (e.g. development of innovative reactive-separation systems).