A machine learning approach to predict strain-specific phage-host interactions
Pamela Y. Camejo, Felipe Suárez Rojas, Alberto de la Ossa, Rodrigo Hurtado, Daniel Tichy, Christian Pieringer, Michael Pino, Paola Mora-Uribe, Soledad Ulloa, Rodrigo Norambuena, Eduardo Tobar-Calfucoy, Matías Aguilera, Victoria Rojas-Martínez, Onix Cifuentes, Andrea Sabag, Nicolas Cifuentes, Daniel San Martín, Claudia Infante, Pablo Cifuentes, Hans Pieringer, Luis León
Abstract
The use of bacteriophages for biological control of bacterial infections is a promising approach to combat antimicrobial resistant bacteria. Prediction of phage-bacteria interactions is key to identify sensitive bacterial strains to phage therapy. Since these interactions are governed by multiple biological mechanisms, it is not a simple task to predict the outcome of a phage infection, which varies even among strains from the same species. In this study, machine learning-based models capable of predicting the host range of phages from sequencing data were developed. Models were trained using phage-bacteria protein-protein interactions (PPI), predicted from PPI databases, and a host-range dataset obtained from experimental assays with 10 Salmonella enterica and 3 Escherichia coli bacteriophages. The performance of prediction models differed among bacteriophages, ranging from 78 to 92% of accuracy in the case of Salmonella and 84-94% in Escherichia phages, with the highest accuracy (94%) achieved for E. coli phage CBDS-07. Results demonstrated the effectiveness of using PPI as a feature to design ML models for phage-bacteria phenotype prediction.