Machine learning methods to identify markers and predict antimicrobial resistance in Escherichia coli

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Antimicrobial resistance (AMR) in pathogenic *Escherichia coli* poses a significant healthcare burden, leading to prolonged hospital stays and increased costs.
• Whole genome sequencing is now a standard tool for analyzing *E. coli* outbreaks and for surveillance.
• In silico methods offer a potentially faster and more cost-effective approach for identifying genomic features related to antimicrobial resistance compared to traditional laboratory methods.

Purpose of the Study:

• To develop and evaluate machine learning (ML) classification models for predicting antimicrobial resistance (AMR) in *E. coli* using whole genome sequencing data.
• To identify novel genomic markers associated with antimicrobial resistance.
• To compare the performance of ML models against established database methods for AMR prediction.

Main Methods:

• Collected and analyzed 4300 *E. coli* whole genome sequences with associated laboratory-derived susceptible, intermediate, or resistant (SIR) data for 34 antimicrobials.
• Trained three ML models—gradient boosted decision trees, support vector machines (SVMs), and artificial neural networks (ANNs)—using 11-length genome subsequences (k-mers).
• Classified isolates as SIR for each antimicrobial using the trained ML models and compared performance with AMRFinderPlus and ResFinder.

Main Results:

• The ML models achieved high average accuracies on the primary dataset: 93.6% (XGBoost), 92.7% (SVM), and 92.8% (ANN), significantly outperforming database methods (AMRFinderPlus: 63.9%, ResFinder: 75.7%).
• On independent datasets, ML models showed strong performance (average accuracies: 81.6% XGB, 79.9% SVM, 81.2% ANN), though ResFinder achieved 94.7% on one dataset.
• ML models demonstrated the capability to identify novel genomic markers of resistance, a key advantage over database approaches.

Conclusions:

• Machine learning models utilizing k-mer analysis of whole genome sequences are effective tools for predicting antimicrobial resistance in *E. coli*.
• These ML models offer a promising, potentially faster and cheaper, alternative to traditional database methods for AMR surveillance and research.
• The ability of ML models to discover novel resistance markers enhances their utility for advancing our understanding and control of antimicrobial resistance.