Integrating bacterial and viral genomic information enhances and extends the machine learning predictions of bacteriophage activity against pathogenic Escherichia coli

The use of bacteriophage (phage) to treat bacterial infections is undergoing a resurgence due to the rise of antibiotic resistance. Pathogenic Escherichia coli, including extraintestinal pathotypes such as uropathogenic E. coli (UPEC), pose a substantial clinical burden with the challenge of multi-drug resistance strains, making it important to progress alternative treatment options such as those based on phage. One major hurdle is selecting effective phage combinations against an infecting strain while accounting for the multiple mechanisms that determine phage susceptibility and bacterial resistance. We previously analysed over 9000 interactions between 31 phage and 314 sequenced E. coli and built machine learning models to predict phage activity against "unseen" E. coli based on these scores and bacterial genome sequence data. However, this approach did not leverage phage gene content, limiting its ability to generalise across phage. Here, we advance this framework by combining bacterial and phage pangenomes into a single predictive model, allowing gene content across related phage as well as bacteria to be incorporated as machine learning features. This unified approach increased predictive accuracy relative to the single-phage models and produced a single model capable of predicting previously unseen phage-E. coli interactions. In leave-phage-out analyses, it also predicted activity for phages whose interaction data were entirely excluded from training. The genomic features most consistently contributing to prediction included bacterial determinants of surface recognition and anti-phage defence, alongside phage-associated features, indicating that this combined model, termed "PanPhage", captures genetic information relevant to multiple stages of the phage-host interaction.