Antimicrobial peptides are a type of peptide capable of exerting antibacterial functions by interacting with bacterial cell membranes or intracellular biomolecules, thereby disrupting bacterial physiological processes and ultimately leading to bacterial death. A novel deep learning model was constructed to screen antimicrobial peptides from soil metagenomic data and validated the screened peptides using techniques such as molecular docking and molecular dynamics simulations. The model demonstrated an outstanding performance with a precision of 98.7%, an accuracy of 96.5%, a recall rate of 91.9%, an F1-score of 95.2%, and a specificity of 99.2%, showcasing excellent efficiency, interpretability, and practical application value alongside robust generalization capabilities. After training, the model successfully identified several short peptides with significant antimicrobial potential, with a subset chosen for further investigation. The findings revealed that the screened peptide Gly-Thr-Ala-Trp-Arg-Trp-His-Tyr-Arg-Ala-Arg-Ser could effectively attach to the bacterial transcription regulator protein MrkH, exhibiting inhibitory effects on Klebsiella pneumoniae, Escherichia coli, and Staphylococcus aureus. This study aimed to provide a theoretical basis for the development and application of new antimicrobials in the food industry by integrating deep learning with molecular simulation technologies.