*Imagine* a patient arrives with a life-threatening Pseudomonas infection resistant to every antibiotic in the hospital formulary. Phage therapy could save them, but which phages from a library of hundreds will actually work against their specific bacterial strain? Traditional testing takes weeks, and the urgency is such that the patient has just days.
The Phage Foundry’s newest preprint addresses this urgency with systematic dataset generation and thorough machine learning (ML) modeling workflow.
The team tackled this challenge by creating the largest systematic P. aeruginosa phage-host interaction dataset to date (9,405 interactions: 95 phages × 99 X/MDR clinical strains) and training ML models that predict phage susceptibility from genome sequences alone.
The results exceeded the team’s expectations. ML models achieved 86% accuracy on completely held-out strains. In head-to-head murine infection experiments, the ML-designed cocktail outperformed an expert-designed cocktail by 12-fold in reducing bacterial burden. The key insight? O-antigen serotype and surface receptor genes dominate infection outcomes, making genome-based prediction feasible.
This establishes a scalable framework for rapidly matching therapeutic phages to patient isolates, collapsing weeks of lab testing into computational predictions that can guide same-day treatment decisions with additional work. While this work focused on P. aeruginosa, the approach is extensible to other pathogens, addressing the full spectrum of drug-resistant infections.
Leave a Reply