Explainable Deep Learning Models Coupled with Peptidomics for Screening Antibacterial Peptides from In Vitro Simulated Digestion Hydrolyzates of Jinhua Ham Broth

An explainable deep learning ensemble model integrated with peptidomics was developed to discover antibacterial peptides generated during simulated gastrointestinal digestion of Jinhua ham broth. Simulated digestion increased small peptides and free amino acids, and the resulting peptides inhibited Escherichia coli , Staphylococcus aureus , Salmonella typhimurium , and Listeria monocytogenes by disrupting cell walls and membranes. Peptidomic analysis identified 1,325, 1,694, and 1,980 peptide sequences in undigested, gastric, and intestinal digests, respectively. A KOA-optimized CNN-BiLSTM-att-RF ensemble was trained on antimicrobial peptide datasets with 1,000–15,000 training entries, and SHapley Additive exPlanations were used to perform explainability analysis of the model. The model trained on 1,3000 entries achieved the best overall performance and was applied to screen intestinal peptides. The ten top-ranked peptides were synthesized, and eight showed minimum inhibitory concentrations (MICs) lower than or not significantly different from that of nisin Z, and the most potent peptide, Pep-1, exhibited an MIC of 7.5 μM.