Automated Annotation and Bayesian Analysis of Metabolic Network Models.

Abstract

Mechanistic models are important tools for understanding, simulating, engineering, and predicting biological systems, but their utility depends on their credibility, and whether they can be interpreted, reused, and analyzed reliably. This dissertation focuses on mechanistic metabolic models, whose use remains limited by challenges in model reproducibility, incomplete annotation, and uncertainty in parameterizing metabolic networks. This dissertation addresses these challenges through the development and evaluation of practices and computational methods for improving metabolic network models. This dissertation begins by examining standards and best practices for reproducible systems biology modeling, emphasizing the role of FAIR principles, model exchange formats, documentation, testing, and public dissemination. It then evaluates Bayesian Metabolic Control Analysis as a framework for estimating metabolic control coefficients from limited physiological data, identifying the types of data most important for accurate inference and clarifying the method’s limitations. The final part of the dissertation presents an AI-assisted pipeline for automated reaction annotation, combining structured biochemical database search with language-model ranking to assign KEGG reaction identifiers to biological models. Together, these studies suggest that scalable and reliable metabolic modeling requires both standardized model representation and improved computational methods for model interpretation and analysis. By examining how models are shared, annotated, and analyzed, this dissertation contributes practical guidance and computational tools for improving the credibility, reusability, and analytical value of metabolic network models.