Fig. 1: Schematic of the study workflow.

a We began with model construction, using bioactive peptide datasets from a previous study. By fine-tuning the pre-trained MP-BERT model, we mainly developed three models, which were ABP-MPB, AFP-MPB, and AOP-MPB, collectively referred to as BAP-MPB (Bioactive peptide for MP-BERT). Furthermore, these three models were used to predict peptides with potential triple activities. b Subsequently, based on the SHAP interpretation of these three models, we applied a 13-AA sliding window to identify KFFs with the highest average SHAP value for each peptide sequence that were predicted to be positive for bioactivity by all three models. Next, with distinct AA features, these KFFs were divided into four subfamilies, and the top three high-frequency AAs at each position were systematically organized into every possible sequence combination to form plausible sequence subspaces. Representative sequences from each plausible sequence subspace were selected as c_AMPs for chemical synthesis. c, d Finally, we conducted further efficacy tests and mechanistic analyses for antimicrobial and antioxidant activities in vitro (n = 3 biologically independent replicates in the MIC assays, mean ± s.d.), as well as in vivo experiments using a mouse sepsis model.