Did AI design an antibiotic that worked in mice?

The model found laboratory-active candidates unusually often

Researchers at Shanghai Jiao Tong University built a two-stage model that represents a peptide sequence together with its attached fatty-acid chain. It first learned antimicrobial patterns from 50,310 linear peptides, then specialised on 477 lipopeptides: 326 active and 151 inactive. In a homology-aware independent test of 61 lipopeptides, it correctly identified 35 of 44 active examples and 14 of 17 inactive ones, with 0.92 precision and 0.80 recall. The team then mined 96,911 bacterial genomes, derived 2,657 short peptide backbones, model-screened their virtual lipid forms and applied additional charge and fatty-acid filters. Of 46 selected candidates synthesized at random, 44 inhibited at least one of 11 laboratory-tested pathogens at 128 micromolar or below. That 95.7% hit rate is the clearest result, although the extra rule-based filters mean it cannot be credited to the neural network alone.[1]

Why it matters—and why this is not a drug

One candidate, LP630, inhibited Gram-positive, Gram-negative and mycobacterial strains and bound several bacterial-membrane components. In two wound experiments, 45 male mice were divided among nine groups of five. Wounds infected with Acinetobacter baumannii or Staphylococcus aureus received six hourly topical doses of LP630, saline or a comparator antibiotic. LP630 reduced bacterial loads by at least tenfold and the higher tested concentrations were associated with smaller wounds during 11 days of follow-up. For people facing antimicrobial-resistant infections, the work shows how computation can narrow an enormous search space before costly synthesis. It does not show safety or effectiveness in humans. The compound lost antibacterial activity in mouse serum, probably through protein binding, which is why the researchers tested a surface wound rather than a bloodstream or systemic infection.[1]

The full evidence record

The model's specialised training base was modest, and its independent lipopeptide test included 44 active and 17 inactive examples. The authors reduced close-sequence overlap with an 80% identity cut-off, compared the chain prompt with sequence-only and other fusion designs, and weighted the minority class. On that test, accuracy was 0.80, Matthews correlation 0.57 and F1 0.85. The prospective synthesis step is more persuasive than benchmark performance because it checks predictions in the laboratory, but the reported positive definition was broad: activity against any tested strain at a minimum inhibitory concentration of 128 micromolar or less. LP630 was stronger against the two strains taken into mice, at 2 and 4 micromolar, yet it was generally less potent in vitro than colistin or daptomycin against their respective bacterial groups.

The animal study used two full-thickness wounds per mouse and five animals per group. It measured colony counts soon after six topical doses, body weight and wound area; it was not a blinded clinical-style comparison and used only young male BALB/c mice. Separate single-dose testing reported no death, weight loss or organ lesions up to 160 milligrams per kilogram over 14 days, but repeated-dose toxicity, pharmacokinetics, resistance evolution and systemic efficacy remain unknown. The authors specifically call for membrane-leakage work and resistance testing. Confidence would rise with independent replication, larger and sex-balanced animal studies, formulation work that overcomes serum binding, longer safety testing, head-to-head dosing against standard care and eventually phased human trials. Funding came from Shanghai municipal, Chinese national research and Shanghai Jiao Tong University programmes; the authors declared no competing interests.[1]