Model-driven engineering of microbial metabolism

Metabolic engineering is a cornerstone of the bio-based economy, which aims to replace fossil-fuel based processes with biological systems. In this thesis, I develop and apply computational methods for model-driven engineering of microbial metabolism. Due to the complexity of biological systems, it is often unclear beforehand exactly which aspects of the systems will be limiting when engineering metabolism. By using the data coming from experiments to inform metabolic models, these models can be used to drive decision-making for the subsequent set of experiments, thus solving obstacles one at a time with the most information possible; a process known as the Design-Build-Test-Learn cycle. Models are thus an important part of metabolic engineering and in this thesis I use several modelling frameworks: constraint-based optimization, flux sampling, as well as kinetic modelling. Each modelling method has its strengths and in chapter 2 I review the use of models in metabolic engineering, consolidating the lessons I learned by applying different modelling methods and working together with experimental partners. We discuss the various questions that should be posed before deciding on a modelling strategy, as well as the strengths and weaknesses of the most common frameworks used for metabolic modelling. In addition, we discuss the different sources of experimental data and their suitability for using them in the different modelling frameworks.

In chapter 3, we apply constraint-based modelling to Pseudomonas putida to support the development of a metabolic valve enabling control over the growth versus production trade-off. Using an existing constraint-based genome-scale model for P. putida, we evaluate the potential of the PDH reaction, producing acetyl-CoA from pyruvate, to serve as this metabolic valve. However, the current model fails to recognize the experimentally confirmed essentiality of this reaction. Thus, we search for alternative pathways in the model that can cause this misprediction and show that with these reactions disabled, the PDH reaction can function as a metabolic valve, allowing growth or the production of pyruvate derived products depending on its activity. This prediction was experimentally verified with the production of pyruvate as well as two pyruvate-derived compounds: 2-ketoisovalerate and lycopene. In chapter 4, we continue our work in P. putida. Using flux sampling, we identify targets for genetic up- and down-regulation for the overproduction of malonyl-CoA, a precursor for the synthesis of fatty acids as well as polyketides. By comparing the simulated fluxes in growing versus producing cells, we find targets that undergo a significant change of flux. These predictions were experimentally verified using a biosensor-based high-throughput screening, as well as through the production of phloroglucinol, a compound derived from malonyl-CoA.

With chapter 5 we turn our attention towards kinetic models. While kinetic models are promising, they are significantly harder to work with due to their increased demands for data, as well as difficulties in simulating and optimizing larger models. Thus, we develop a pipeline for the automated generation of these types of models and explore how well this automated approach holds up with missing or inaccurate estimates of the kinetic parameters. Finally, we investigate the use of automated model-reduction methods, which we further explore in chapter 6. Here, we focus on the topic of automated model reduction, as the scale at which constraint-based models can be applied is much larger than the applicable scale for kinetic models. We investigate how well several model reduction methods, operating on the constraint-based model, reproduce the behaviour of the full-sized kinetic model after estimating the parameters with a time-series data set. We find that these reduced models can be representative of the full-scale model and can thus be a useful proxy for engineering the whole system.

In chapter 7, we return to P. putida for the production of curcumin, but now with the aid of a kinetic model. Curcumin is a yellow-coloured compound, originally isolated from the turmeric plant, used as a food additive and colouring agent. Furthermore, the precursors for the production of curcuminoids can be derived from lignin, a substantial component of plant waste biomass, making it an interesting target for the bio-based, circular economy. The curcumin production pathway utilizes only a few enzymes but can lead to several products depending on enzyme kinetics and relative concentrations of intermediates. This makes it an interesting test case for a kinetic model, as constraint-based models do not model these details due to their steady-state assumption. Through sensitivity analysis on the kinetic model, we perform optimal experimental design to predict the experiments most informative for improving the model used to study the pathway.

Finally, in chapter 8 I discuss how these different chapters contributed to the objectives of this thesis, and look towards how to further integrate computational models of metabolism into metabolic engineering studies in the future. In addition, I elaborate on challenges and future opportunities for using kinetic metabolic models.