Transcriptome-Based Machine Learning Models to Predict Antimicrobial Resistance in Pseudomonas aeruginosa

Abstract

Antimicrobial resistance (AMR) is a growing threat in planetary health and demands innovative systems biology strategies for rapid and accurate detection of AMR and attendant resistance phenotypes. Chief among the AMR cases is Pseudomonas aeruginosa that exhibits remarkable genomic adaptability and contributes to multidrug resistance. This study aimed to evaluate the potential of transcriptome-based machine learning (ML) models to predict AMR in P. aeruginosa and attendant gene expression signatures. We integrated transcriptomic profiles of clinical isolates (n = 414) with ML algorithms to predict resistance to four antibiotics: ceftazidime, ciprofloxacin, meropenem, and tobramycin. ML models achieved high predictive accuracy, with the tobramycin model attaining 98.8% accuracy and 100% sensitivity. Each of the four antibiotics yielded distinct transcriptomic signatures enriched in pathways such as biofilm formation, membrane transport, virulence, and amino acid metabolism. Importantly, 10 gene signatures were identified across all four antibiotics, implicating them in core resistance mechanisms including oxidative stress response and iron acquisition. We further identified a core set of 10 mRNAs that are consistently deregulated in resistant isolates across all four drugs, pointing to a shared transcriptional program underpinning multidrug resistance. In conclusion, the transcriptome-based signatures reported herein (1) provide promising candidates for translational research toward development of mechanism-guided diagnostic assays for AMR in P. aeruginosa, and (2) attest to the potential of transcriptome-based ML models to predict AMR. Further studies and validation in independent cohorts are called for.