Supplementary Datasets - Comprehensive interaction profiling and machine learning prediction of bacteriophage infectivity across clinically diverse Pseudomonas aeruginosa

AbstractThe rise of antibiotic-resistant bacterial infections has driven renewed interest in bacteriophage therapy, where viruses that specifically kill bacteria are used as targeted antimicrobials. Pseudomonas aeruginosa, a WHO critical-priority pathogen that causes severe infections in hospitalized and immunocompromised patients, presents a major challenge for phage therapy because of its extraordinary genetic diversity. Phages effective against one bacterial strain often fail against others, and existing cross-resistance-profiling approaches require iterative empirical testing of each new patient isolate. To establish a genome-based framework for rapid phage-isolate matching, we assembled a collection of 95 genomically diverse P. aeruginosa phages representing 20 genera and tested each against 99 genetically diverse clinical isolates, generating 9,405 infection outcome measurements. Bacterial O-antigen serotype emerged as the dominant determinant of strain susceptibility, while defense systems, anti-defense systems, and prophage burden contributed smaller strain-specific effects. The full curated multivariate model explained 47% of strain-susceptibility variance. Machine-learning models integrating these features and pangenome-derived gene clusters reached a per-strain AUROC of 0.86. In an in vivo proof-of-concept test against a single held-out strain, the ML-designed cocktail produced a ~12-fold greater median CFU reduction than the expert-designed cocktail (q = 0.045), with both cocktails substantially reducing burden relative to the untreated control (~113-fold for ML, ~9-fold for CG; both q < 10⁻³). SHAP analysis of the model identified bacterial surface-architecture genes (LPS biosynthesis, outer membrane proteins, type IV pili) as the dominant predictors, with defense-system content modulating which specific phages succeed against a strain rather than uniformly damping susceptibility. Together, these results establish a genome-based framework for predicting phage susceptibility in genetically diverse clinical isolates. Dataset S1. Per-strain prophage counts stratified by CheckV completeness threshold (99 strains). Dataset S2. Full 95-phage × 99-strain interaction matrix (9,405 scored interactions). Dataset S3. Per-fold, per-strain, and per-phage AUROC/AUPR/MCC for all models, with paired-test statistics. Dataset S4. Per-tercile model performance across four difficulty axes (phylogenetic novelty, phenotypic divergence, marginal difficulty, full-matrix degree). Dataset S5. Full quartile × variant × top-k cocktail-success table (144 rows; backs Table S14). Dataset S6. Per-strain counts of 118 defense-system families. Dataset S7. Predicted O-antigen serotype and OSA coverage per strain. Dataset S8. Per-prophage geNomad predictions (coordinates, completeness, taxonomy). Dataset S9. Prophage counts by viral realm and per-strain gene counts.