Abstract
Antimicrobial resistance (AMR) in Pseudomonas aeruginosa poses a critical global health challenge, with current diagnostics relying on slow, culture-based methods. Here, we present a machine learning (ML) framework leveraging transcriptomic data to predict antibiotic resistance with high accuracy. We applied a genetic algorithm to 414 clinical isolates to identify minimal, highly predictive gene sets, each consisting of approximately 35 to 40 genes, that distinguish resistant from susceptible strains for meropenem, ciprofloxacin, tobramycin, and ceftazidime. Automated ML classifiers trained on these sets achieved accuracies of 96-99% on held-out test data (macro-average F1 scores of 0.93-0.99), surpassing clinical deployment thresholds. Interestingly, multiple distinct, non-overlapping gene subsets exhibited comparable predictive performance, indicating that resistance acquisition broadly impacts the expression of diverse regulatory and metabolic genes. Comparison with known resistance markers from CARD and operon annotations revealed a substantial number of previously unannotated gene clusters, highlighting significant knowledge gaps in current AMR understanding. Mapping these genes onto independently modulated gene sets (iModulons) revealed transcriptional adaptations occurring across diverse genetic regions. Overall, this study presents a streamlined machine-learning workflow for transcriptomic data and offers a pathway toward rapid diagnostics and personalized treatment strategies against AMR.
Competing Interest Statement
The authors have declared no competing interest.
Funder Information Declared
National Institutes of HealthNational Institutes of Health, https://ror.org/01cwqze88, 5R35GM143009
Nebraska Ethanol Board AwardNebraska Ethanol Board Award, , 26-1106-0157-001
University of NizwaUniversity of Nizwa, https://ror.org/01pxe3r04,