A generative AI model designed a batch of antimicrobial peptides from scratch. Ten were made and tested. Eight killed bacteria. One of them, named Arcinin, cleared the hardest drug-resistant pathogens in a dish and then healed infected wounds in mice.
The work, published July 7 in Nature Communications ↗, is a rare case of computer-designed peptides carried the whole distance from an algorithm to a live animal. Most AI drug-design papers stop at a docking score or a single test-tube assay. This one ran the gauntlet.
The system is called ARCADIAMP. At its core is a diffusion model, the same family of generative AI that powers image generators, retooled to produce amino acid sequences instead of pixels. The model proposes candidate peptides, and a second AI, a classifier built on a protein language model, scores how likely each one is to actually kill bacteria. The two loop against each other, the generator learning to make sequences the classifier rates highly, iteration after iteration, before anything touches a lab bench.
What a peptide antibiotic has to be
Antimicrobial peptides are short chains of amino acids that most living things make as a first line of defense. They work by physically punching holes in bacterial membranes rather than gumming up a specific enzyme, which is why bacteria have a harder time evolving resistance to them. That is also their problem. The same membrane-wrecking trick can shred human cells, they fall apart in blood, and the sequence space is so vast that finding a good one by hand is close to hopeless.
So the bar is not just potency. A usable peptide has to kill bacteria at low concentrations, leave human cells alone, and survive in serum long enough to matter. ARCADIAMP was built to optimize all three at once, and Arcinin is the payoff.
Against the ESKAPE pathogens, the six drug-resistant bacteria that cause most hospital infections and that the World Health Organization flags as priority threats, Arcinin worked at 8 to 32 micrograms per milliliter, a low and clinically useful concentration. It barely touched human red blood cells (the concentration needed to burst half of them was above 512 micrograms per milliliter, more than sixteen times the dose that kills bacteria). And it held activity in 50 percent blood serum against four of the ESKAPE species, the stability check that sinks most peptide candidates.
The researchers then watched it work. Electron microscopy, membrane-voltage assays, time-kill curves, and molecular dynamics simulations all pointed to the expected mechanism: Arcinin inserts into and punches through the bacterial membrane in under a microsecond. In a mouse with an infected wound, it cut the bacterial load by about 10,000-fold (a 4-log reduction) and let the skin re-grow and heal.
Design versus dredging
This is the second distinct route to new antimicrobial peptides worth watching. The other is mining nature at scale: earlier this year a team pulled 873 candidate peptides out of Arctic deep-sea metagenomes ↗, reading sequences that already exist in the wild. ARCADIAMP does the opposite. It invents sequences that were never in any genome, steered toward the properties you want rather than the ones evolution happened to leave lying around.
Both approaches feed the same shortage. The antibiotic pipeline has been thin for decades because the economics are bad and the chemistry is hard, while resistant infections keep climbing. Antimicrobial peptides are one of the few classes where the biology still offers open ground, and peptidemodel hosts a growing set of antimicrobial peptide cards ↗ for exactly that reason.
The caution is the usual one for a preclinical result. A 4-log kill in a mouse wound is a strong signal, not a drug. Arcinin has not been through toxicity studies at scale, pharmacokinetics, or anything resembling a human trial, and the graveyard of antimicrobial peptides that looked great in mice is well populated. What is new here is not the molecule. It is that a generative model produced a selective, serum-stable peptide that held up all the way to an animal, on an eight-in-ten hit rate, which is the part that scales.