Open access
Author
Date
2023Type
- Doctoral Thesis
ETH Bibliography
yes
Altmetrics
Abstract
Metabolism is the chemistry of small molecules enacted by living systems. The main functions of metabolism are: the supply of chemical driving force to cellular processes, the generation of precursors for biological macromolecules, and transport. The rate at which a metabolic reaction proceeds is called flux. What fluxes result from the confluence of evironmental and organismal factors is a question of fundamental interest. Fluxes are intimately related to cellular growth, thus making them an important phenotype in applications such as the design of bio-production processes or the study of cancer. For reactions that happen inside of a cell, fluxes are not directly measurable and must therefore be inferred through a statistical model. In the introductory chapter of this thesis we establish the concepts necessary to build such a model for metabolic flux inference. We decided to interpret our model in the Bayesian paradigm. Doing so forces us to model our prior knowledge about fluxes as a probability distribution. In chapter 2, we introduce three distinct probability distributions, each one models a different aspect of our prior knowledge and is useful in distinct application scenarios. The data we use for metabolic flux inference consists of the relative concentrations of metabolites that differ only in their isotopic composition, called mass-isotopomers. In chapter 3 we develop a model that describes the data-generating process of a liquid-chromatography mass-spectrometry method that is used to measure mass-isotopomers. We calibrate this model using a data-set of over 400 measurements and compare our model to an observation model that has been used as the gold standard in literature on flux inference. We find that the assumptions of our model significantly change our posterior beliefs about fluxes compared to the model from literature when both are conditioned on the same observations. In chapter 4 we designed two Monte-Carlo algorithms to draw samples from the Bayesian posterior distributions that represent our beliefs about fluxes upon observing data. In the common scenario where we would like to infer fluxes for many experiments with a limited computational budget, the Monte-Carlo algorithms do not suffice. Therefore, we develop a machine-learning approach that relies on neural spline flows that is amenable to high-throughput analyses. For this approach to function, we introduced a novel cylinder-embedding for fluxes and a log-ratio transormation for mass-isotopomers. Using machine-learning for flux inference opens up opportunities to study them with higher precision and in higher throughput than previously possible. Show more
Permanent link
https://doi.org/10.3929/ethz-b-000655552Publication status
publishedExternal links
Search print copy at ETH Library
Publisher
ETH ZurichSubject
Simulation based inference; Bayesian statistics; metabolic flux analysisOrganisational unit
08839 - Zamboni, Nicola (Tit.-Prof.)
More
Show all metadata
ETH Bibliography
yes
Altmetrics