Abstract
Genome-scale metabolic models (GEMs) are indispensable tools for probing cellular metabolism, enabling predictions of metabolic fluxes, guiding strain optimization, and advancing biomedical research. However, their predictive capacity is often compromised by incomplete reaction networks, stemming from gaps in biochemical knowledge, annotation inaccuracies, and insufficient experimental validations. Here we present MuSHIN (Multi-way SMILES-based Hypergraph Interface Network), a deep hypergraph learning method that integrates network topology with biochemical domain knowledge to predict missing reactions in GEMs. Evaluated on 926 high- and intermediate-quality GEMs with artificially removed reactions, MuSHIN achieves up to a 17% improvement over the current state-of-the-art method across multiple evaluation metrics. Furthermore, MuSHIN substantially enhances phenotypic predictions in 24 draft GEMs associated with fermentation by resolving critical metabolic gaps, as validated against experimental measurements. Together, these findings highlight MuSHIN's potential to advance GEM reconstruction and accelerate discoveries in systems biology, metabolic engineering, and precision medicine.