BIONEWS
A New Search for Antibiotics: Inside the Proteins Linked to Brain Disease
Penn researchers use AI to find hidden antibiotic peptides — dubbed 'prionins' — tucked inside prion and prion like proteins.
Penn researchers use AI to find hidden antibiotic peptides — dubbed 'prionins' — tucked inside prion and prion like proteins.
Researchers at the University of Pennsylvania used a deep-learning platform to scan millions of short protein fragments derived from nearly 3,000 prion and prion-like proteins, a family best known for misfolding in neurodegenerative disease, and surfaced more than 1,000 candidate antimicrobial peptides, a class they named prionins (Genetic Engineering & Biotechnology News summary of the Penn study). Fifty-nine showed activity in lab tests; two reduced the burden of a drug-resistant hospital pathogen in mice.
The story reads, on the surface, like another entry in the "AI discovers a drug" column. That framing buries the actual finding. The model the Penn group trained, APEX 1.1, is a capable sequence classifier, but it is not, on its own, what made the study work. What made it work is the upstream human choice to point that classifier at prion and prion-like proteins: a family of about 3,000 sequences whose reputation in biology comes from misfolding, aggregation, and diseases like Creutzfeldt-Jakob and Alzheimer's, not from fighting infection. Ask a similar model to search the usual antimicrobial peptide libraries, and it returns the usual hits. The non-obvious decision was to widen the search universe, not to sharpen the search instrument.
That distinction matters for how readers should weight this result. The work, led by César de la Fuente's Machine Biology Group at Penn, is a proof of concept that useful antimicrobial activity can hide in protein families the field had classified as off-limits for this kind of inquiry. It is not a clinical breakthrough, and the caveats are real. Of the more than 1,000 candidate prionins the model flagged, 59 were synthesized and tested in bacterial assays; only two were advanced to mice, and in those mice they reduced the burden of Acinetobacter baumannii, a drug-resistant hospital pathogen that public-health agencies have flagged as a priority for new antibiotic development. No human data exists. The path from a mouse-active peptide to a usable drug is long, expensive, and prone to failure, and the "prionin" label itself is author-coined, a marketing choice for a class of molecules that still has to earn its keep in the lab.
The bigger lesson is structural. Most of the antibiotics in clinical use were discovered by screening natural products: soil bacteria, fungi, marine organisms, a method that has returned diminishing returns for decades. The Penn approach is a different kind of search. Take a deep-learning classifier, point it at a protein family no one has asked about in this context, and let the model do the triage. The model doesn't make the discovery. The scientist makes the discovery by deciding where the model is allowed to look. In an era of increasingly capable AI biology tools, that boundary, meaning which proteomes, which taxa, which "dark" sequence space, is where the leverage actually sits.
The next test is whether the prionin class survives contact with the rest of drug development: stability in blood, toxicity in animals, manufacturing at scale, and activity against the broader set of resistant pathogens the field needs to address. If even a small fraction of the prionin list clears preclinical filtering, the result will not be that AI found a drug. It will be that a human used AI to ask an old molecule a new question, and the molecule answered.