Unraveling principles of thermodynamics for genome-scale metabolic networks using graph neural networks.

Our understanding of metabolic thermodynamics is limited by the lack of genome-scale data on the standard Gibbs free energy change (Δ_rG°) of metabolic reactions. Here, we present dGbyG, a graph neural network (GNN)-based model for predicting Δ_rG° with superior accuracy, versatility, robustness, and generalization ability. Integration of dGbyG predictions into metabolic networks facilitated model curation, improved flux prediction accuracy, and identified thermodynamic driver reactions (TDRs) with substantial negative values of the reaction Gibbs free energy change (Δ_rG). TDRs showed distinctive network topological features and heterogeneous enzyme expression, implying coupling between reaction thermodynamics and network topology for efficient metabolic regulation. We also discovered a universal pattern of thermodynamics in linear metabolic pathways, explained by a multi-objective optimization model balancing the needs to maximize pathway flux and minimize enzyme and metabolite loads. Our work expands accessible thermodynamic data and elucidates optimality principles in metabolism at the genome scale. A record of this paper's transparent peer review process is included in the supplemental information.