Machine learning lists almost a million possible antibiotic peptides from the world’s microbes

The atlas

Microbes make small peptides to fight each other, but they are hard to spot: the genes are short and easy to miss in DNA data. The team trained machine-learning models to find them and ran them over more than 60,000 metagenomes, DNA collected from soil, oceans, guts and other habitats, plus tens of thousands of microbial genomes.

The result, called AMPSphere, holds about 860,000 candidate antimicrobial peptides, almost all previously unknown. It is free to search.

The test

The researchers made 100 of the candidates. Seventy-nine were active against bacteria in some way and 63 killed at least one pathogen of concern, including drug-resistant strains. A few were tested in mice with skin infections and reduced bacterial load.

The data also show where the peptides come from. Many sit in parts of genomes that move between species, which hints at how microbes trade weapons.

Why it is useful

The atlas is less a list of drugs than a map of where to look. It gives other labs a starting set and a way to see which kinds of peptides recur across habitats. Turning any of them into a medicine will need the slow work of fixing toxicity, stability and manufacturing that has kept most antimicrobial peptides out of the clinic.

Source Santos-Júnior C. D. et al., Cell, June 2024 (AMPSphere)

Amino Weekly reports research. It is not medical advice and never gives doses.