Prediction of Antibiotic Resistance Phenotypes and Minimum Inhibitory Concentrations in Salmonella Using Machine...

Traditional experimental methods for determining antibiotic resistance phenotypes (ARPs) and minimum inhibitory concentrations (MICs) in bacteria are laborious and time consuming. This study aims to explore the potential of whole-genome sequencing data combined with machine learning models for robustly predicting ARPs and MICs in Salmonella. Using a training set of 6394 Salmonella genomes alongside antimicrobial susceptibility testing results, we built two machine learning (ML) predictive models based on the pan-genome and pan-resistome. Each model was implemented using three algorithms: random forest, extreme gradient boosting (XGB), and convolutional neural network. Among them, XGB achieved the highest overall accuracy, with the pan-genome and pan-resistome models accurately predicting ARPs (98.51% and 97.77%) and MICs (81.42% and 78.99%) for 15 commonly used antibiotics. Feature extraction from pan-genome and pan-resistome data effectively reduced computational complexity and significantly decreased computation time. Notably, fewer than 10 key genomic features, often linked to known resistance or mobile genes, were sufficient for robust predictions for each antibiotic. This study also identified challenges, including imbalanced resistance classes and imprecise MIC measurements, which impacted prediction accuracy. These findings highlight the importance of using multiple evaluation metrics to assess model performance comprehensively. Overall, our findings demonstrated that ML, utilizing pan-genome or pan-resistome features, was highly effective in predicting antibiotic resistance and identifying correlated genetic features in Salmonella. This approach holds great potential to supplement conventional culture-based methods for routine surveillance of antibiotic-resistant bacteria.