Antimicrobial Resistance Prediction in PATRIC and RAST

Area of Science:

• Genomics
• Computational Biology
• Infectious Diseases

Background:

• Antimicrobial resistance (AMR) poses a significant global health threat due to evolving resistance mechanisms and a declining pipeline of effective antibiotics.
• Accurate prediction of bacterial pathogen resistance phenotypes before culturing is crucial for timely clinical decision-making and effective treatment strategies.

Purpose of the Study:

• To advance the prediction of antimicrobial resistance (AMR) phenotypes and identify genomic correlates of resistance in bacterial pathogens.
• To leverage curated genomic data and metadata to build predictive models for AMR.

Main Methods:

• Curated a comprehensive database of bacterial genomes with associated AMR metadata, including minimum inhibitory concentrations, on the PATRIC FTP server.
• Developed custom AdaBoost (adaptive boosting) machine learning classifiers to predict specific AMR phenotypes.
• Validated classifiers for resistance in Acinetobacter baumannii, Staphylococcus aureus, Streptococcus pneumoniae, and Mycobacterium tuberculosis.

Main Results:

• Achieved high accuracies (88-99%) for predicting carbapenem resistance in A. baumannii, methicillin resistance in S. aureus, and beta-lactam/co-trimoxazole resistance in S. pneumoniae.
• Attained accuracies ranging from 71-88% for predicting resistance to isoniazid, kanamycin, ofloxacin, rifampicin, and streptomycin in M. tuberculosis.
• Integrated these classifiers into the RAST and PATRIC annotation services for species-specific AMR phenotype and genomic feature prediction.

Conclusions:

• Machine learning models effectively predict bacterial antimicrobial resistance phenotypes using genomic data.
• The developed framework provides a valuable tool for species-specific AMR prediction, aiding in clinical and research applications.
• This approach supports faster identification of resistance mechanisms, crucial for managing the global AMR crisis.