Consensus approach for assessing and resolving uncertainty in genome-scale metabolic models: advancing systems-level understanding of microbial metabolism

Genome-scale metabolic models (GEMs) are an important methodology in systems biology: they play a major role in investigating microbial metabolism and predicting responses to perturbations by representing bacterial metabolism as a whole system. GEMs can be automatically reconstructed from bacterial genomes with computational tools that employ distinct methodological approaches. My initial analyses of GEMs reconstructed by different tools for a small set of diverse gut microbial species revealed that automated reconstruction pipelines often produce GEMs with different structures and predictive behaviour, even for the same organism. Because individual models can excel at different tasks and capture distinct metabolic capabilities, I hypothesised that combining them can increase confidence in network content and improve performance. In this thesis, I present a consensus approach implemented in GEMsembler, a Python package for cross-tool model comparison and assembly of consensus models from any subset of input GEMs. Alongside the consensus strategy, GEMsembler offers broad analysis functionality, including detection and visualisation of biosynthetic pathways, growth evaluation, and an agreement-driven curation workflow. In a use-case study, consensus models curated with this workflow, combining four automatically reconstructed GEMs for Lactiplantibacillus plantarum and Escherichia coli, outperform gold-standard models on auxotrophy and gene-essentiality benchmarks. Moreover, enabled by GEMs comparison, optimising gene–protein–reaction (GPR) rule combinations derived from input and consensus models improves gene-essentiality predictions even for manually curated gold-standard models. GEMsembler also helps explain model performance by highlighting relevant metabolic pathways and, together with the consensus principle, supports the assessment of network uncertainty and informs the design of targeted experiments to resolve it. Finally, I apply the consensus approach to de novo reconstruction of two of the most abundant human gut bacteria, Bacteroides uniformis and Phocaeicola vulgatus, yielding first-iteration curated models that reproduce their growth and major metabolic phenotypes. In agreement with experimental data, the B. uniformis model secretes more lactate and malate but grows less, whereas the P. vulgatus model secretes less of these metabolites, grows more, and also secretes succinate. Together, these results show that the consensus approach facilitates building metabolic models that are more accurate, concise, and biologically informed, advancing systems-level understanding of microbial metabolism.