Machine learning–driven prediction of antimicrobial resistance in Staphylococcus aureus using genomic and allelic signatures
Abstract
IntroductionAntimicrobial resistance (AMR) in Staphylococcus aureus is a growing clinical threat, and machine learning (ML) offers a scalable approach to identifying allele-level resistance determinants.MethodsUsing a dataset of 300 clinical isolates, we trained and compared seven ML models - LSTM, CNN, Multi-Layer Perceptron (MLP), Decision Tree, XGBoost, Logistic Regression, and Stochastic Gradient Descent (SGD) – to predict resistance across six clinically important antibiotics from allele, locus-tag, and EggNOG-annotation features. Training-set class imbalance was corrected with SMOTE and all seven models were retrained under one identical pipeline.ResultsXGBoost achieved the strongest overall test-set performance (accuracy 95%, ROC AUC 0.9745, F1-score 0.9474), closely matched by Decision Tree (accuracy 95%, F1-score 0.9520). Both outperformed the deep-learning architectures (LSTM, CNN, MLP), which showed pronounced overfitting after class-imbalance correction. SHapley Additive exPlanations (SHAP) analysis identified “cat__allele_Cluster_1015_Allele_8” as the most influential feature; this association is hypothesis-generating and reflects statistical importance within this dataset, not a confirmed causal or biological mechanism.DiscussionThe genomic feature set was selected once, on the full isolate collection, by a prior SVM-RFE pipeline before the present train/test split, and no external, temporal, or multi-center validation dataset was used. Reported performance should therefore be read as an upper-bound, best-case estimate rather than a validated benchmark. These findings support ML, and ensemble methods such as XGBoost and Decision Tree in particular, as candidate components of future AMR surveillance and diagnostic pipelines, pending external and prospective validation.
Related
- Staphylococcus aureus in Bovine Mastitis: Pathogenesis, Antimicrobial Resistance, and Emerging Control Strategies
- Staphylococcus aureus in bovine mastitis in northwestern Iran: patterns of antimicrobial and multidrug resistance, biofilm formation, and virulence/resistance gene carriage
- Staphylococcus pettenkoferi : an underrecognized species with clinically relevant antimicrobial resistance. A comprehensive genomic European study
- Methicillin resistance and antimicrobial non-susceptibility trends among Staphylococcus aureus clinical isolates in Aseer, Saudi Arabia (2013–2024): A retrospective laboratory-based surveillance study
- Artificial Intelligence in the Fight Against Antimicrobial Resistance: Applications in Diagnosis, Susceptibility Prediction, Precision Dosing, and Antibiotic Discovery.
BibTeX
@article{ghosh2026machine,
title = {{Machine learning–driven prediction of antimicrobial resistance in Staphylococcus aureus using genomic and allelic signatures}},
author = {Joyeta Ghosh and Ravi Kant and Jyoti Taneja},
journal = {Frontiers in Microbiology},
year = {2026},
month = nov,
doi = {10.3389/fmicb.2026.1877073},
url = {https://www.frontiersin.org/articles/10.3389/fmicb.2026.1877073},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Machine learning–driven prediction of antimicrobial resistance in Staphylococcus aureus using genomic and allelic signatures
AU - Joyeta Ghosh
AU - Ravi Kant
AU - Jyoti Taneja
PY - 2026
DA - 2026/11/04
JO - Frontiers in Microbiology
DO - 10.3389/fmicb.2026.1877073
UR - https://www.frontiersin.org/articles/10.3389/fmicb.2026.1877073
AB - IntroductionAntimicrobial resistance (AMR) in Staphylococcus aureus is a growing clinical threat, and machine learning (ML) offers a scalable approach to identifying allele-level resistance determinants.MethodsUsing a dataset of 300 clinical isolates, we trained and compared seven ML models - LSTM, CNN, Multi-Layer Perceptron (MLP), Decision Tree, XGBoost, Logistic Regression, and Stochastic Gradient Descent (SGD) – to predict resistance across six clinically important antibiotics from allele, locus-tag, and EggNOG-annotation features. Training-set class imbalance was corrected with SMOTE and all seven models were retrained under one identical pipeline.ResultsXGBoost achieved the strongest overall test-set performance (accuracy 95%, ROC AUC 0.9745, F1-score 0.9474), closely matched by Decision Tree (accuracy 95%, F1-score 0.9520). Both outperformed the deep-learning architectures (LSTM, CNN, MLP), which showed pronounced overfitting after class-imbalance correction. SHapley Additive exPlanations (SHAP) analysis identified “cat__allele_Cluster_1015_Allele_8” as the most influential feature; this association is hypothesis-generating and reflects statistical importance within this dataset, not a confirmed causal or biological mechanism.DiscussionThe genomic feature set was selected once, on the full isolate collection, by a prior SVM-RFE pipeline before the present train/test split, and no external, temporal, or multi-center validation dataset was used. Reported performance should therefore be read as an upper-bound, best-case estimate rather than a validated benchmark. These findings support ML, and ensemble methods such as XGBoost and Decision Tree in particular, as candidate components of future AMR surveillance and diagnostic pipelines, pending external and prospective validation.
SN - 1664-302X
ER -