Researchers have developed a machine learning system that predicts which viruses can infect bacterial strains using genome sequences alone.
Published in Nature Microbiology on September 29, the study addresses a problem in phage research: finding a virus that works against the specific bacterium being targeted.
Bacteriophages, or phages, infect bacteria. Their selectivity makes them attractive candidates for tackling resistant infections, but also means that selecting useful combinations requires extensive testing.
The team developed its approach using six datasets containing 115,037 interactions across 949 bacterial strains and 518 phages. Experimental validation assessed 1,240 interactions involving previously unseen phages and E. coli strains.
The model achieved an AUROC of 0.84 in that validation, a measure of how well it distinguishes interacting pairs from noninteracting ones.
In computational tests, selecting five phages with the model covered up to 97.5% of bacterial strains in the E. coli dataset. That figure describes predicted selection evaluated against interaction data, not treatment success in patients.
Performance varied across datasets, and generalizing to unfamiliar bacterial genera remained difficult. The advance is a way to prioritize promising laboratory candidates.