Redundancy-aware AI-guided discovery and experimental validation of antimicrobial peptides for infected wound treatment.

Abstract

Artificial intelligence (AI) is accelerating antimicrobial peptide (AMP) discovery, but prediction-centered workflows often overlook dataset redundancy, peptide synthesizability, and experimental anti-infective translation. Here, we developed a redundancy-aware AI-guided peptide discovery workflow integrating redundancy-controlled dataset construction, model interpretation, candidate screening, synthesis-linked experimental validation, and evaluation in an infected-wound model. A redundancy-retention (RR) dataset of 1861 peptides with E. coli MIC annotations was compared with CD-HIT-filtered CD60-CD90 datasets containing 439-1061 sequences. Redundancy control reshaped activity-density distributions, SHAP-derived feature dependence, and virtual-screening stringency. Screening 2.1 million random 13-mer peptides yielded 1763, 189, 157, 141, and 20 candidates from CD60, CD70, CD80, CD90, and RR workflows, respectively. Experimental synthesis and MIC testing showed that the RR-derived group had a higher mean crude yield and a higher hit rate against E. coli than the CD90-derived group (60% vs 20%; MIC ≤16 μM). The lead peptide A36 showed broad activity against the tested Gram-negative bacteria, inhibited drug-resistant clinical A. baumannii isolates, displayed low hemolysis and cytotoxicity, retained substantial integrity in serum and antibacterial activity after protease exposure, disrupted bacterial membranes, and reduced bacterial burden while promoting wound closure in an A. baumannii-infected wound model. These findings identify redundancy control as a practical factor influencing candidate selection, synthetic accessibility, and experimental hit recovery in AI-guided AMP discovery.