NEW Preprint: A comprehensive phage-bacteria interaction atlas links phage lineage and capsule serotype to genome-guided machine learning prediction in Klebsiella pneumoniae

In our newest preprint, the Phage Foundry tackles a hard question: how do we fight superbugs when our strongest antibiotics fail? 

When a WHO critical-priority pathogen like Klebsiella pneumoniae becomes resistant to drugs, doctors run out of options fast. Viruses that eat bacteria (called phages) can help, but finding the right one normally takes weeks in a lab, time that a sick patient simply does not have.

The Phage Foundry tackled this urgency with systematic dataset generation and a genome-guided ML modeling workflow. The team built one of the most comprehensive K. pneumoniae phage–host interaction datasets to date, a fully crossed atlas of 8,484 interactions testing 84 taxonomically diverse phages against 101 globally sourced clinical strains, including some of the extreme multidrug-resistant isolates.

Trained on these datasets, a genome-only model (GenoPHI) predicted phage susceptibility for strains it had never seen and recovered the capsule biology that governs Klebsiella phage infection, without being told where to look. It proved most valuable on atypical strains, the ones where simple serotype matching breaks down. The team then used the model to design phage cocktails directly from genome sequence and then validated them in vitro and in a gnotobiotic mouse model.

This work establishes a scalable framework for matching therapeutic phages to patient isolates, turning weeks of lab screening into computational predictions that can guide treatment decisions with additional validation. It’s the second pathogen the team has mapped this way, following their P. aeruginosa study, and part of a broader effort to build the data foundations that genome-guided phage therapy will require.