Machine learning-based prediction of multi-level antimicrobial resistance in Klebsiella pneumoniae using whole-genome sequencing data.

OBJECTIVES: In this study, we developed an integrated approach for accurate and comprehensive prediction of anti-microbial resistance (AMR) using whole-genome sequencing (WGS) data. METHODS: We developed machine learning (ML) models using WGS data from 5239 strains, spanning three independent, geographically and temporally diverse cohorts. The models were designed for multi-level AMR prediction, including resistant/susceptible (R/S), resistant/intermediate/susceptible (R/I/S), and high/low-level resistant (H/L) status against 11 antibiotics. We utilized nine ML algorithms and three anti-microbial susceptibility testing interpretation standards (CLSI, EUCAST, ECOFF), with robust five-fold cross-validation. RESULTS: The model for distinguishing R/S categories exhibited excellent discrimination capability with area under the receiver operating characteristic curve (AUC) for all 11 antibiotics > 0.9 and a mean categorical agreement (CA) of 0.96. It also demonstrated robust performance across diverse regions (including Europe, the Americas, and Asia), sequence types, isolation sources, and over a long time span (2004-2022). We further developed models to simultaneously predict resistance (R), intermediate (I), and susceptible (S), as well as high and low levels of antibiotic resistance, achieving a mean AUC and CA > 0.9. The mean very major error, major error, and minor error for R/I/S models were 0.013, 0.015, and 0.024, respectively, indicating great promise for clinical application. CONCLUSIONS: MLWA, our proposed model, leverages the largest dataset to date to enable accurate, multi-level AMR predictions that are generalizable across regions and stable over time. This ML-WGS approach enhances genotype-to-phenotype interpretation and facilitates precise and rapid anti-microbial selection, representing a significant step forward in managing AMR.