Guiding the Refinement of Biochemical Knowledgebases with Ensembles of Metabolic Networks and Machine Learning

Mechanistic models explicitly represent hypothesized biological knowledge. As such, they offer more generalizability than data-driven models. However, identifying model curation efforts that improve performance for mechanistic models is nontrivial. Here, we develop a solution to this problem for genome-scale metabolic models. We generate an ensemble of models, each equally consistent with experimental data, then perform simulations with them. We apply machine learning to the simulation output to identify model structure variation that maximally influences simulations. These variants are high-priority candidates for curation through removal, addition, or reannotation in the model. We apply this approach, automated metabolic model ensemble-driven elimination of uncertainty with statistical learning (AMMEDEUS), to 29 bacterial species to improve gene essentiality predictions. We explore targets for individual species and compile pan-species targets to improve the database used during model construction. AMMEDEUS is an automated and performance-driven recommendation system that complements intuition during curation of biochemical knowledgebases.