Oyarzun, Diego

Mac Aodha, Oisin

Merzbacher, Charlotte

UKRI AI Centre for Doctoral Training in Biomedical Innovation

2026-01-22T13:34:50Z

2025-12-02

Metabolism is the set of all biochemical reactions that sustain life. Recent advances in genetic engineering have enabled the design of metabolic circuits which produce chemicals of interest in microbial hosts, including pharmaceuticals, cosmetics, and biofuels. However, designing new metabolic pathways and understanding complex interactions in metabolism is challenging, especially when considering systems which operate across multiple scales of cellular organization, such as gene regulation, signalling pathways and cellular metabolism. Computational mechanistic models can speed up the development of microbial strains for chemical production and improve our understanding of disease states, and are widely used across the study of metabolism. There are many classes of mechanistic models, including ordinary differential equations and genome-scale linear models, but all can be slow and challenging to construct due to the large amounts of domain-specific knowledge they require. Recent work in machine learning has developed a wide range of tools which can detect patterns in and make highly accurate predictions from large data sets without require many assumptions about biological mechanism. Many of these methods, however, require a large amount of expensive or unavailable data. This PhD thesis presents three approaches which aim to bridge the gap between these two modelling paradigms. Firstly, I present a method for efficient optimization of ordinary differential equation (ODE) models of biological circuits across multiple temporal and spatial scales. The method relies on Bayesian Optimization, a technique commonly used to fine-tune deep neural networks, to learn the shape of a performance landscape and iteratively navigate the design space towards an optimal circuit. This strategy allows the joint optimization of both circuit architecture and parameters, and hence provides a feasible approach to solve a highly non-convex optimization problem in a mixed-integer input space. Second, I integrate ODE models of pathways with genome-scale metabolic models of the production host to create a new joint simulator method, which combines fine-grained concentration trajectory prediction with information about the dynamic global state of native metabolism. I implement machine learning surrogate models to enable accelerated simulation. Finally, I employ machine learning models to predict gene deletion phenotypes from synthetic data generated from genome-scale mechanistic models. I achieve state-of-the-art accuracy for gene essentiality in various organisms, as well as the first predictive model for small molecule production from deletion screening data. Each of these results demonstrates a way machine learning can be used to improve mechanistic models: to optimize model structure, to replace slow computation, and to improve predictive accuracy.

https://era.ed.ac.uk/handle/1842/44336

https://doi.org/10.7488/era/6856

en

The University of Edinburgh

Cain, S., C. Merzbacher, and D. A. Oyarzun (2024). “Low-dimensional representations of genome-scale metabolism”. In: Foundations of Systems Biology in Engineering Conference, pp. 2024–05

Merzbacher, C. (2022). “A Machine Learning Approach for Optimization of Gene Circuits for Metabolic Engineering”. MA thesis. University of Edinburgh

Merzbacher, C. and D. A. Oyarzún (2023). “Applications of artificial intelligence and machine learning in dynamic pathway engineering”. In: Biochemical Society Transactions 51.5, pp. 1871–1879

Merzbacher, C., O. Mac Aodha, and D. A. Oyarzún (2023). “Bayesian optimization for design of multiscale biological circuits”. In: ACS Synthetic Biology 12.7, pp. 2073– 2082.

Merzbacher, C., O. M. Aodha, and D. A. Oyarzún (2025). “Modelling dynamic host-pathway interactions at the genome scale”. In: Metabolic Engineering.

machine learning

mechanistic models

metabolism

metabolic engineering

Mechanistic models and machine learning for metabolism

Thesis or Dissertation

Doctoral

PhD Doctor of Philosophy