A deep-learning model read 19.3 million protein fragments cut from prion proteins and flagged 1,179 that might kill bacteria. Prions are the proteins behind mad cow disease and Creutzfeldt-Jakob, infamous for misfolding into clumps that wreck the brain. A group at the University of Pennsylvania asked the opposite question of them: not what makes them dangerous, but whether short stretches buried inside their sequences could work as antibiotics. The model said yes, more than a thousand times over. The team named the hits prionins.
The work, a brief communication in Nature Microbiology ↗ from Cesar de la Fuente-Nunez's Machine Biology Group, treats a protein's amino acid sequence as a library to be searched. The researchers pulled 2,897 prion and prion-like proteins from the UniProt database, chopped them into every possible fragment 8 to 50 amino acids long (19,324,138 of them), and ran the lot through APEX, a neural network trained to predict whether a peptide will stop bacteria from growing. The 1,179 fragments it ranked as likely active came from 1,068 different organisms, spanning fungi, animals, and single-celled life. Encrypted antibiotic potential is scattered all across prion sequence space, not concentrated in one odd protein.
Predictions are cheap. The test was synthesis. The team made 75 of the top candidates, each picked to be less than 70 percent similar to any known antimicrobial peptide and to each other, then put them against 11 disease-causing strains. Fifty-nine of the 75 stopped at least one strain, and 42 did it at low concentrations (a minimum inhibitory concentration of 16 micromolar or less, meaning a small dose was enough to halt growth). Most of the activity landed on Gram-negative bacteria, the ones wrapped in an extra outer membrane that makes them the hardest to kill and the reason the World Health Organization ranks species like Acinetobacter baumannii among its top drug-resistant threats.
How they work looks familiar. Prionins punch holes in bacterial membranes: in the lab they made the outer membrane leaky and collapsed the voltage across the inner one, often as hard as the reference antibiotics polymyxin B and levofloxacin run alongside them. Many were shapeless in water and folded into ordered structures only on contact with a membrane, the same trick natural host-defense peptides like the human molecule LL-37 ↗ use. Toxicity, the usual reason these molecules die in development, was low: only 1 of the 75 ruptured human red blood cells, and 16 showed neither that nor toxicity to human kidney cells at the highest doses tested.
Two went into mice. Prionin-7, drawn from a fungus, and prionin-38, from the roundworm Caenorhabditis elegans, were dripped onto Acinetobacter baumannii skin abscesses one hour after infection at ten times their active dose. By day four, prionin-38 had cut the bacterial load roughly a thousandfold, matching polymyxin B, while prionin-7 managed about a tenth of that. No mouse lost weight. Peptides predicted from sequence alone, with no natural role anyone has observed, cleared a real infection.
That last point is where the authors are careful, and worth repeating. The paper does not claim prionins are released during infection or that they serve as immune defenses in the organisms they came from. It shows that prion-related proteins are a productive place to go looking for antibiotic leads, not that evolution put the antibiotics there on purpose. The honest framing matters when the headline writes itself.
The method is the throughline. This is the same lab that has mined the human body's own proteins, extinct species, and the global microbiome for encrypted antibiotics, and peptidemodel has covered two recent runs of the playbook: a generative model that rewrote 100 antimicrobial drafts ↗ and an Arctic deep-sea sweep that surfaced 873 candidates ↗. What is new here is the source. If antibiotics can be hiding inside the proteins most associated with disease, the search space for the next antimicrobial peptide ↗ is larger and stranger than the field assumed.