Large-scale genomic analysis reveals significant role of insertion sequences in antimicrobial resistance of Acinetobacter baumannii

Acinetobacter baumannii, a prominent nosocomial pathogen renowned for its extensive resistance to antimicrobial agents, poses a significant challenge in the accurate prediction of antimicrobial resistance (AMR) from genomic data. Despite thorough researches on the molecular mechanisms of AMR, gaps remain in our understanding of key contributors. This study utilized rule-based and three machine learning models to predict AMR phenotypes, aiming to decipher key genomic factors associated with AMR. Genomes and antibiotic resistance phenotypes from 1,012 public isolates were employed for model construction and training. To validate the models, a data set comprising 164 self-collected strains underwent next-generation sequencing, nanopore long-read sequencing, and antimicrobial susceptibility testing using the broth dilution method. It was found that the presence of antibiotic resistance genes (ARGs) alone was insufficient to accurately predict AMR phenotype for the majority of antibiotics (90%, 18 out of 20) in the public data set. Conversely, it was observed that combining ARGs with insertion sequence (IS) elements significantly enhanced predictive perform ance. The Random Forest model was found to outperform the support vector machine (SVM), logistic regression model, and rule-based method across all 20 antibiotics, with accuracies ranging from 83.80% to 97.70%. In the validation data set, even higher accuracies were achieved, ranging from 85.63% to 99.31%. Furthermore, conserved sequence patterns between IS elements and ARGs were validated using self-collected long-read sequencing data, substantially enhancing the accuracy of AMR prediction in A. baumannii. This study underscores the pivotal role of IS elements in AMR.