Genome-based machine learning for predicting antimicrobial resistance in Salmonella isolated from chicken

Antimicrobial resistance (AMR) continues to pose a grave threat to public health. The increase in the burden of AMR is fueled by the indiscriminate use of antimicrobial agents in agriculture. The objective of this study was to develop a genome-based machine learning model to predict AMR in Salmonella isolated from chicken meat. Genomic information on 205 Salmonella isolates from chicken samples was combined with the AMR phenotype data of these isolates to amoxicillin-clavulanic acid, ampicillin, ceftiofur, ceftriaxone, sulfisoxazole, streptomycin, tetracycline, and cefoxitin. Four machine learning algorithms i.e., logit boost, random forest, support vector machine, and extreme gradient boosting were trained on this data to build models. The best-performing model for each antimicrobial was used to predict the AMR phenotypes of a new set of 200 Salmonella isolates also from chicken, and the predictions were compared to AMR phenotype predictions from ResFinder. The machine learning models showed high sensitivity (≥ 0.833), specificity (≥ 0.833), and balanced accuracy (≥ 0.866), across all the tested antimicrobials. The models predicted resistance prevalences ranging from 1% (ceftriaxone) to 65.5% (streptomycin). When the AMR phenotype predictions of the machine learning models were compared to predictions from ResFinder, the predictions from this study were accurate (> 95%).