A Closed-Loop Multimodal AI Framework for Discovering Surface-Active Antimicrobial Peptides

Abstract

The activity of antimicrobial peptides (AMPs) is typically characterized in solution, yet their translation into infection-resistant implant coatings requires covalent immobilization onto surfaces, a fundamentally different physicochemical context for which predictive models and application-specific data are critically lacking. Here, we introduced a closed-loop, AI-driven discovery framework that integrated three complementary peptide generation strategies with a multimodal deep-learning classifier to predict AMP activity directly in the surface-tethered state. Trained on only 122 experimentally characterized surface-immobilized sequences, the model achieved an area under the curve of 0.989 and an F1 score of 0.941 through the synergistic fusion of physicochemical descriptors, sequence-engineering features, BioBERT semantic embeddings, and a raw-sequence convolutional stream. From a generated library of 788933 unique sequences, 15 candidates spanning high and low surface activity predictions were selected for experimental validation. Covalent grafting onto titanium surfaces under kinetically controlled conditions enabled antimicrobial assays that revealed strong concordance between computational predictions and measured antibacterial efficacy against P. aeruginosa, where high-activity candidates showed mean bacterial reductions of 82.9–91.3% across strategies, substantially exceeding the low-activity controls at 38.6–50.5%. This work established a transferable blueprint for the data-efficient discovery of surface-active antimicrobial peptides, closing the loop between computational design and functional biomaterial validation.