Multidrug-resistant (MDR) bacteria are important public health problems. Antibiotic susceptibility testing (AST) currently uses time-consuming culture-based procedures, which cause treatment delays and increased mortality. We developed a machine learning model using Acinetobacter baumannii as an example to explore a fast AST approach using metagenomic next-generation sequencing (mNGS) data. The key genetic characteristics associated with antimicrobial resistance (AMR) were selected through a least absolute shrinkage and selection operator (LASSO) regression model based on 1,942 A. baumannii genomes. The mNGS-AST prediction model was accordingly established, validated, and optimized using read simulation sequences of clinical isolates. Clinical specimens were collected to evaluate the performance of the model retrospectively and prospectively. We identified 20, 31, 24, and 3 AMR signatures of A. baumannii for imipenem, ceftazidime, cefepime, and ciprofloxacin, respectively. Four mNGS-AST models had a positive predictive value (PPV) greater than 0.97 for 230 retrospective samples, with negative predictive values (NPVs) of 100% (imipenem), 86.67% (ceftazidime), 86.67% (cefepime), and 90.91% (ciprofloxacin). Our method classified antibacterial phenotypes with an accuracy of 97.65% for imipenem, 96.57% for ceftazidime, 97.64% for cefepime, and 98.36% for ciprofloxacin. The average reporting time of mNGS-based AST was 19.1 h, in contrast to the 63.3 h for culture-based AST, thus yielding a significant reduction of 44.3 h. mNGS-AST prediction results coincided 100% with the phenotypic AST results when testing 50 prospective samples. The mNGS-based model could be used as a rapid genotypic AST approach to identify A. baumannii and predict resistance and susceptibility to antibacterials and could be applicable to other pathogens and facilitate rational antimicrobial usage.
Keywords: Acinetobacter baumannii; antimicrobial resistance; clinical metagenomics; machine learning; rapid antimicrobial susceptibility testing.