Machine learning of metabolite-protein interactions from model-derived metabolic phenotypes.

Unraveling metabolite-protein interactions is key to identifying the mechanisms by which metabolism affects the function of other cellular layers. Despite extensive experimental and computational efforts to identify the regulatory roles of metabolites in interaction with proteins, it remains challenging to achieve a genome-scale coverage of these interactions. Here, we leverage established gold standards for metabolite-protein interactions to train supervised classifiers using features derived from genome-scale metabolic models and matched data on protein abundance and reaction fluxes to distinguish interacting from non-interacting pairs. Through a comprehensive comparative study, we explore the impact of different features and assess the effect of gold standards for non-interacting pairs on the performance of the classifiers. Using data sets from Escherichia coli and Saccharomyces cerevisiae, we demonstrate that the features constructed by integrating fluxomic and proteomic data with metabolic phenotypes predicted from genome-scale metabolic models can be effectively used to train classifiers, accurately predicting metabolite-protein interactions in the context of metabolism. Our results reveal that the high performance of classifiers trained on these features is unaffected by the method used to generate gold standards for non-interacting pairs. Overall, our study introduces valuable features that improve the performance of identifying metabolite-protein interactions in the context of metabolism.