Mechanism-Guided Antimicrobial Peptide Design through Membrane-Surface Fingerprinting and Graph Diffusion

Abstract

Antimicrobial peptides (AMPs) are promising antibiotic alternatives, but current AI-based discovery often relies on sequence labels and lacks explicit modeling of peptide-membrane interactions. We developed Membrane-MaSIF, an MD-derived membrane surface matching model designed to capture peptide-membrane compatibility beyond sequence-level AMP labels. Using LL-37-perturbed Acinetobacter baumannii outer membranes, Membrane-MaSIF represents binding, insertion, and pore-like perturbation states as computable surface fingerprints encoding local geometry and physicochemical features. We further integrated Membrane-MaSIF with PepGraph-Diffusion for de novo AMP candidate generation and applied it to an external SPLUNC1 α4-derived A4 analogue library for known-scaffold prioritization. Experimental validation of selected candidates, including peptide 71 and high-scoring A4 analogues, supported the utility of membrane surface matching for enriching membrane-active antibacterial peptides. This framework provides a mechanism-aware strategy for Gram-negative AMP discovery and optimization.