Utilizing whole genome sequencing data for machine learning driven prediction of antibiotic resistance in Escherichia coli

Le résumé fourni par la source

Background Antimicrobial resistance (AMR) is an escalating global public health threat that substantially undermines the effectiveness of standard anti-infective therapies and increases the risk of adverse clinical outcomes. Methods We analyzed whole-genome sequencing (WGS) data from 1952 Escherichia coli isolates with AST phenotypes. Gene, single-nucleotide polymorphisms (SNPs), and k-mer features were extracted to train machine-learning classifiers for predicting resistance to 10 common antibiotics. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) and other classification metrics. Results Across all antibiotics and feature representations, model AUCs ranged from 0.6691 to 0.9879. Gene-based and integrated-feature models showed superior and stable performance (AUC 0.8936–0.9787 and 0.8888–0.9879, respectively) compared with SNP-based models and k-mer models. Prediction performance was near-saturated for aminoglycosides (GEN and TOB) and ciprofloxacin (CIP), whereas β-lactams exhibited greater heterogeneity, with amoxicillin/clavulanate (AMX/CLA) exhibiting the lowest AUC. Feature-importance analysis highlighted 40 core genes that were highly concordant with established resistance mechanisms, including aac(3)-IIg / aac(3)-IId for aminoglycosides, blaTEM , blaSHV-12 , blaCTX-M-14 , and blaOXA for β-lactams, and tetR(A) / tet(A) for tetracyclines. Conclusion This study demonstrated that machine learning models built on WGS data can accurately and efficiently predict resistance phenotypes to 10 commonly used antibiotics in Escherichia coli . Among the evaluated feature representations, gene-based and integrated multi-feature approaches yielded the most robust and reliable performance across antibiotic-specific tasks, highlighting the practical utility of WGS-derived genomic features for rapid AMR phenotype prediction and future clinical decision-support applications.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.