Phenotypic antibiotic resistance prediction using antibiotic resistance genes and machine learning models in Mannheimia haemolytica

Mannheimia haemolytica is one of the most common causative agents of bovine respiratory disease (BRD); however, antibiotic resistance in this species is increasing, making treatment more difficult. Integrative-conjugative elements (ICE), a subset of mobile genetic elements (MGE), encoding up to 100 genes have been reported in Mannheimia haemolytica genomes to confer multidrug resistance, including resistance to antibiotics commonly used in the treatment of BRD. However, the presence of antibiotic resistance genes (ARGs) does not always agree with phenotypic resistance. Prior investigations have reported an overall phenotype-genotype concordance less than 75 % in BRD pathogens. The objective of the current study was to compare genotype-phenotype concordance either by annotating known resistance genes in genomes or predicting antibiotic resistance determinants de novo with machine learning (ML). The genotype-phenotype concordance rates of ARGs were generally > 90 %, while that of ML models were > 80 %. For all seven antibiotics (danofloxacin, enrofloxacin, florfenicol, tetracycline, tildipirosin, tilmicosin, and tulathromycin), the genotype-phenotype concordance rates were more accurate with ARGs. The annotations of ML models for all antibiotics included various types of sequences, including coding sequences such as DNA topoisomerase IV (danofloxacin) and non-coding sequences near tetracycline genes (multiple antibiotics), MGE (tetracycline and tildipirosin), or virulence genes (danofloxacin and enrofloxacin). When tested on an external set of isolates for validation, the best predictor of antibiotic resistance performed similarly to the training/testing datasets for each antibiotic. Incorporating single nucleotide polymorphisms and ARGs unknown during previous studies resulted in higher concordance rates (>90 %) for fluoroquinolones and tilmicosin, respectively. By finding increased concordance rates for known ARGs, this study was able to show that ARGs should continue to be utilized to predict phenotypic resistance.