SMILES-Based Deep Learning Enables Discovery of Chemically Modified Ultrashort Antimicrobial Peptides

Abstract

Abstract Sequence-based models for antimicrobial peptides are limited in representing chemically modified peptides, restricting computational discovery to canonical amino acids. Here, we developed SMAMP, a SMILES-based graph deep learning framework that encodes peptides as molecular graphs and captures noncanonical residues and chemical modifications. SMAMP accurately predicted Escherichia coli MIC (MSE = 0.558, R2 = 0.542, PCC = 0.741), outperforming another SMILES-based baseline. Guided by SMAMP, we designed 19 ultrashort peptides incorporating terminal modifications or noncanonical amino acids, 13 of which exhibited antibacterial activity. Predicted MICs correlated with experimental measurements (r = 0.764). Lead peptides HP-1 and HP-2 showed potent bactericidal activity, low mammalian cytotoxicity, and reduced resistance development compared with conventional antibiotics. Both peptides protected Galleria mellonella from infection and promoted wound healing in mice. Proteomic and mechanistic analyses revealed membrane disruption coupled with oxidative stress. SMAMP provides a chemically informed strategy for discovery and optimization of modified antimicrobial peptides.