Machine learning for identifying antibacterial agents derived from yam (Dioscorea) and elucidating antibacterial mechanisms based on metabolomics

Abstract

Saponins from edible plants are promising natural antibacterial agents because of their structural diversity, abundance, and safety. To rapidly identify antibacterial saponins, this study developed a machine learning-based screening strategy using multimodal early fusion of 1D&2D, 3D, and MACCS molecular features. Predictive models were built and systematically evaluated, and the fusion model showed excellent performance on an independent test set. Guided by this model, hecogenin, a potent antibacterial compound from yam, was identified. In a lettuce-based food application system, hecogenin significantly reduced bacterial contamination. Metabolomic analysis showed that it inhibited Staphylococcus aureus by disrupting tryptophan, phenylalanine, and tyrosine metabolism. Molecular simulations further indicated that its antibacterial activity was linked to binding the key enzymes AroA and AroB in the amino acid biosynthesis pathway. This study provides an effective strategy for discovering antibacterial agents from edible plants and highlights the potential of dietary saponins as food preservatives.