Efficient computational methods for sampling-based metabolic flux analysis

Abstract

The aim of metabolic flux analysis is to determine the rates at which the processes in metabolism take place. Stationary isotopomer labeling experiments are the state-of-the-art method to generate data for metabolic flux analysis. The analysis of such experiments requires an atom transition model which is able to simulate the carbon atom transitions that take place in metabolism. The operational state of metabolism is represented by the rates at which the considered processes take place. We call this operational state the flux distribution, and it is a parameter of the atom transition model. By comparing the results of the model simulation against experimental data, we gain information about the flux distribution. To increase the identifiability of this inverse problem, we use constraint-based modeling, i.e. we restrict the flux distribution by applying linear constraints that can be derived directly from the stoichiometry of the considered processes. We took a probabilistic view on this inverse problem. We developed computational methods for the complete computational pipeline which is required to carry out metabolic flux analysis based on stationary isotopomer labeling experiments. First, we developed methods for the parametrization of the solution space that arises from constraint-based modeling. We then implemented the software necessary to simulate and evaluate data from labeling experiments. We next formulated the probabilistic framework which describes labeling experiments. The key to carrying out this probabilistic analysis was the development of efficient sampling methods that are able to sample from polytope-supported probability distributions in high dimensions. We first improved the efficiency of existing MCMC methods for sampling uniformly from convex polytopes. We then developed an efficient sampling procedure for the sampling of general convex polytopes-supported probability distribution based on nested sampling. We analyzed datasets from labeling experiments and compared different methods for the computation of confidence intervals for the estimated fluxes. We further generated synthetic data representing simulated labeling experiments, outlining new ways of experimental design.