Researchers have found antibiotic-like peptides hiding inside prion proteins, the biological shipwrecks we usually blame for fatal brain disease.

Call me old-fashioned, but when a protein has a reputation for turning brains into rough seas, I do not usually expect it to be carrying medical treasure in the hold. Prions and prion-like proteins are famous for misfolding, aggregating, and generally behaving like a crew that ignored every warning from the lighthouse. Yet Torres, Wan, and de la Fuente-Nunez just sailed into that fog with deep learning and came back with “prionins,” short antimicrobial peptides tucked inside prion-related proteins like contraband in a barrel of salted cod.

Their paper in Nature Microbiology screened 19.3 million peptide fragments from 2,897 curated prion-related proteins using APEX 1.1, a deep learning system built for antimicrobial peptide discovery. The model flagged 1,179 candidate peptides. Then the team synthesized 75 of them and tested the lot against bacterial pathogens. Fifty-nine inhibited at least one pathogen. Fifty-three appeared to disturb bacterial membranes. Two reduced Acinetobacter baumannii infection burden in mice.

That is not a full medicine chest yet. But it is a mighty interesting glint on the horizon.

A prion is a misfolded protein that can coax similar proteins into the wrong shape, leading to amyloid aggregates and severe neurodegenerative disease. That is the grim sea story most of us know. Antimicrobial peptides, meanwhile, are short chains of amino acids that many organisms use as part of innate defense. They often work by damaging bacterial membranes, which is the cellular equivalent of springing a leak below deck.

The twist here is that amyloid-like sequences are not always villains. Some amyloids do useful jobs, and earlier studies have linked amyloid-associated proteins, including amyloid-beta and cellular prion protein, to antimicrobial or host-protective activity. Biology, the old trickster, rarely respects our tidy filing cabinets.

So the researchers asked a sharp question: what if prion-related proteins contain encrypted antimicrobial fragments? Not whole proteins acting as antibiotics, mind you, but smaller peptide pieces hidden inside larger sequences. Like finding a dagger inside a walking stick. Unsettling, but handy in a storm.

Deep Learning as the Lookout in the Crow’s Nest

Searching 19.3 million fragments by hand would be the kind of task given to a cursed intern in a maritime horror novel. Instead, the team used APEX 1.1 to scan peptides between 8 and 50 amino acids long, predicting which ones might inhibit pathogens at useful concentrations.

This is where machine learning earns its rations. Antimicrobial peptides have patterns: charge, hydrophobicity, length, shape, and other properties that influence whether they can latch onto and disrupt bacterial membranes. Deep learning models can spot combinations that are too slippery for simple rules. They are not magic compasses. They are more like experienced navigators who have seen enough reefs to guess where the next one might be.

The prionins were not just copies of known antimicrobial peptides, either. The authors selected experimentally tested candidates with less than 70% similarity to known AMPs and to each other. In plain terms: they tried not to haul aboard the same fish twice.

Why This Matters When the Antibiotic Seas Are Getting Ugly

Antimicrobial resistance is not some distant squall. The WHO estimates that bacterial antimicrobial resistance directly caused 1.27 million deaths in 2019 and was associated with 4.95 million deaths. We need new antibiotic starting points, and the usual discovery routes have been coming back with too many empty nets.

Peptides are tempting because they can attack bacteria in ways that differ from classic antibiotics, often by disturbing membranes rather than blocking one tidy molecular target. That may make resistance harder in some cases, though no captain worth his salt would claim bacteria cannot adapt. Bacteria have been surviving hostile waters for billions of years. They do not panic easily.

This study joins a growing flotilla of AI-assisted peptide discovery work. Recent reviews argue that machine learning can help with AMP identification and design, while papers on molecular de-extinction and explainable peptide optimization show models mining extinct proteomes, oral microbiomes, and virtual sequence space for new candidates. The prionin paper adds a stranger chart: disease-associated protein families may hide useful antimicrobial cargo.

Reef Warnings Before Anyone Opens the Champagne Barrel

The authors are careful about what they have and have not shown. They did not prove that prionins naturally get released during infection. They did not prove these peptides act as innate immune effectors in their original organisms. And a peptide that works in vitro, or even helps in a mouse infection model, still has to survive the long voyage through toxicity, stability, dosing, manufacturing, immune effects, and pharmacokinetics.

Many a promising molecule has foundered on those rocks. I have seen models overfit, peptides clump, assays flatter, and drug candidates vanish into the fog like a poorly insured schooner.

Still, the hit rate here is hard to ignore: 59 active peptides out of 75 tested. That suggests the model was not merely pointing at random driftwood. It found a real patch of biological ocean worth charting.

The Treasure Is the Search Strategy

The biggest lesson may not be “prions are antibiotics now.” They are not. The better lesson is that biological sequence space still has locked compartments, and deep learning can help pry them open.

Prion-related proteins were an odd place to look, which is precisely why the result has bite. If antimicrobial peptides can hide there, perhaps other protein families we think we understand are carrying useful fragments too. Drug discovery often advances when someone asks a question that sounds slightly unhinged at first light.

So raise a lantern for this crew. They sailed into the prion fog, kept the sails trimmed, and came back with a new class of peptide leads. The harbor is still far off, but the chart just got more interesting.

References

1. Torres, M. D. T., Wan, F. & de la Fuente-Nunez, C. “Deep learning reveals antimicrobial peptides within prions.” Nature Microbiology (2026). DOI: 10.1038/s41564-026-02408-1. PMID: 42321536

2. Wan, F., Wong, F., Collins, J. J. & de la Fuente-Nunez, C. “Machine learning for antimicrobial peptide identification and design.” Nature Reviews Bioengineering 2, 392-407 (2024). DOI: 10.1038/s44222-024-00152-x. PMCID: PMC11756916

3. Wang, B. et al. “Explainable deep learning and virtual evolution identifies antimicrobial peptides with activity against multidrug-resistant human pathogens.” Nature Microbiology (2025). DOI: 10.1038/s41564-024-01907-3. PMID: 39825096

4. Wan, F., Torres, M. D. T., Peng, J. et al. “Deep-learning-enabled antibiotic discovery through molecular de-extinction.” Nature Biomedical Engineering 8, 854-871 (2024). DOI: 10.1038/s41551-024-01201-x. PMID: 38862735

5. “Antimicrobial resistance.” World Health Organization fact sheet. https://www.who.int/news-room/fact-sheets/detail/antimicrobial-resistance

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.