AINEWS
A Quiet Redefinition of 'Prion' Is Why an AI Just Found New Antibiotic Candidates
Prions, the misfolded proteins behind mad cow disease, were reclassified in the 2010s to include a much larger family of proteins.
Prions, the misfolded proteins behind mad cow disease, were reclassified in the 2010s to include a much larger family of proteins.
For decades, "prion" meant one thing in medicine: a misfolded protein that templates the misfolding of its neighbors, slowly destroying the brain in diseases like mad cow and Creutzfeldt-Jakob. The category was small, the prognosis was universally fatal, and the biology was treated as a closed book of horrors.
Then, mostly in the 2010s, a separate community of cell biologists began publishing evidence that the same misfolding-and-templating behavior showed up across a much larger family of proteins involved in ordinary cellular housekeeping: proteins that build stress granules, organize the inside of the nucleus, and help embryos develop. The work, which drew on concepts from liquid-liquid phase separation and low-complexity protein domains, gradually reclassified the prion concept from a narrow pathology label into a broad functional category. The original neurodegenerative diseases kept their clinical horror; the term "prion" quietly grew to cover a much larger source space.
That reclassification is what made a recent University of Pennsylvania result possible. In a paper published in Nature Microbiology, a team reported using a deep-learning model to screen prions and prion-like proteins for hidden antimicrobial peptides, short protein fragments capable of killing bacteria. The screen surfaced several dozen candidate peptides with predicted antibacterial activity, and two lead candidates already showed efficacy treating bacterial infections in mouse models, according to a Gizmodo summary of the work.
The interesting part is the order of operations. The AI did not discover a new biological phenomenon. It screened a category that human researchers had spent roughly a decade expanding, applying a computational filter to a source space that biologists had just learned to see. A screen run against the old, narrow prion category would have surfaced very few candidates. Run against the new, broad one, it surfaced several dozen. The definitional move came first; the computational result inherited it.
The result itself is genuine. The two lead candidates treated skin infections in mice, a standard early-stage test for antibiotic candidates. Antimicrobial resistance is a real and growing problem, and any productive new source space for antibiotics is worth attention. But the work is preclinical: the candidates have not been tested in humans, the full mechanism is still being mapped, and the same misfolding behavior that makes these proteins interesting also underlies the diseases that made the prion concept notorious in the first place. Mining the category for antibiotics is not the same as neutralizing the disease.
What to watch is the next layer. If the reclassification holds up under continued scrutiny, the prion-like protein universe becomes a standing candidate pool for computational screens of this kind, and the actors who control those screens, including the labs with the curated datasets, the model architectures, and the funding to run them at scale, become the actual gatekeepers of which fragments of biology ever become drugs. The University of Pennsylvania result is a useful first data point. The more important question is who else is now running similar screens against the same expanded source space, and what they choose to publish when they do.