Graph deep learning guides the design of chemically modified ultrashort antimicrobial peptides

Summary. Researchers at Zhejiang University of Technology in Hangzhou, China, reported in the Journal of Medicinal Chemistry, published online on 7 October, a deep learning framework called SMAMP that represents antimicrobial peptides as molecular graphs built from their full chemical structure rather than as strings of standard amino acids. Because the model reads every atom, it can handle noncanonical residues and terminal modifications that sequence-based models cannot represent. The authors report that SMAMP predicted the minimum inhibitory concentration against Escherichia coli more accurately than another structure-based baseline. Guided by the model, the team designed 19 ultrashort peptides carrying terminal modifications or noncanonical amino acids, and 13 of them showed antibacterial activity, with predicted and measured values correlating at r = 0.764.

Research context (RUO). The paper reports that two lead peptides, HP-1 and HP-2, killed bacteria in laboratory tests, showed low toxicity toward mammalian cells in culture and led to less resistance development than conventional antibiotics in the authors' assays. Both were also evaluated in a wax moth larva (Galleria mellonella) infection model and in a mouse wound model; these are preclinical results, not data from people. Proteomic and mechanistic analyses pointed to membrane disruption combined with oxidative stress. For laboratories the main interest is the method: a model that scores chemically modified peptides directly could shorten design, synthesis and testing cycles for short peptides that sit outside the 20 standard amino acids. Testing against other bacterial species will show how well the predictions carry over.