Machine learning helped find 860,000 potential peptide antibiotics in nature

Australian, American, Irish, Spanish, Chinese and German researchers have applied machine learning algorithms to analyse available genomic and metagenomic databases to predict and catalogue possible antimicrobial peptides. These are short sequences (10-100 amino acid residues) that inhibit the growth of certain bacteria in some way (most commonly by disrupting the microbial wall). Luis Pedro Coelho (Luis Pedro Coelho) from Queensland University of Technology and colleagues from five countries analysed a dataset of 63410 metagenomes and 87920 genomes of prokaryotes (both free-living and associated with multicellular host organisms) using machine learning. They eventually compiled the AMPSphere catalogue, which included 863498 non-redundant potential antimicrobial peptides, most of which are not available in existing databases. A report of the work is published in the journal Cell.

As a proof of concept, the researchers synthesised and tested in vitro and in vivo 100 predicted peptides potentially active against clinically relevant pathogens as well as commensals of the human gut microbiota. Of these, 79 showed antimicrobial activity, with 63 acting specifically on highly pathogenic antibiotic-resistant ESKAPEE bacteria. In experiments on mice with skin abscesses, their action resembled the effect of polymyxin B, a peptide antibiotic used in clinical practice.