Do Genes Tell the Whole Story? Comparing Genetic Clues and Wet-Lab Tests for Antibiotic Resistance in Salmonella

Abstract

Foodborne salmonellosis remains a major global health concern, and the rise of antimicrobial-resistant (AMR) Salmonella strains poses a challenge to both public health and food safety. Whole-genome sequencing (WGS) is a promising alternative to phenotypic antimicrobial susceptibility testing (AST), yet the performance of WGS-based prediction tools requires thorough evaluation across antimicrobial classes. This study assessed the concordance between phenotypic AST and genotypic predictions generated by two WGS pipelines, abritAMR and Staramr. For the analysis, 33 poultry-associated Salmonella enterica isolates were tested against 15 antibiotics spanning seven classes. Antimicrobial resistance determinants were identified in silico and compared with minimum inhibitory concentration-based AST results. Across antibiotics, both pipelines demonstrated strong and comparable performance, achieving similar positive predictive values (abritAMR 83%; Staramr 80%) and negative predictive values (abritAMR 93%; Staramr 95%). High sensitivity and specificity were observed for most antibiotics, including amikacin, ceftriaxone, chloramphenicol, ciprofloxacin, gentamicin, and streptomycin, where both pipelines reached 100% sensitivity with specificity ranging from 88% to 100%. Cohen’s κ statistics revealed strong agreement (κ = 0.81–1) for several antibiotics in both abritAMR and Staramr. However, substantial variability was observed for several β-lactams and trimethoprim–sulfamethoxazole, where sensitivity ranged widely (20–100%, depending on the drug and pipeline), reflecting predictive limitations for these classes. Both pipelines met the minimum acceptable performance criteria for 10/15 antibiotics evaluated. Notably, abritAMR demonstrated better overall performance for kanamycin, cefoxitin, and ceftiofur, achieving higher sensitivity and specificity (80–100%) and agreement (κ = 0.81). The disagreements between phenotype and genotype observed may be attributed to factors like gene expression variability, differences in database content, and the complexity of genotype–phenotype relationships. Addressing these gaps may require refinement of pipeline parameters and drug-specific adjustments to improve predictions. Overall, these findings indicate that WGS-based AMR prediction is suitable for several antibiotics, particularly aminoglycosides, fluoroquinolones, and phenicols.