Phenotypic antibiotic resistance prediction in Mannheimia haemolytica isolates from cattle with bovine respiratory disease using resistance databases and machine learning

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. Treatment for BRD involves a variety of antimicrobials, with initial treatment failing for about 30% of BRD cases and supporting the idea of increasing antimicrobial resistance. Currently the best way to test for phenotypic resistance is antibiotic susceptibility testing (AST) in a laboratory, however the process requires many days making more immediate treatment difficult. If a nucleic acid assay could be designed using known antibiotic resistance genes (ARGs) and their presence/absence in Mannheimia haemolytica for a more immediate assessment of antimicrobial resistance. A previous study (Owen et al., 2017 https://doi.org/10.1534/g3.117.1137) however found low phenotype-genotype concordance for ARGs associated with antibiotics used to treat BRD. In this study, we employed a machine learning algorithm to identify novel genomic sequences with enhanced accuracy in predicting antibiotic resistance, specifically targeting those commonly used to treat BRD, compared to known ARGs. Mannheimia haemolytica bacterial samples (n = 87) were collected from animal disease diagnostic laboratories. These samples were isolated from lung tissue or nasal swabs of cattle showing clinical signs of BRD. Minimum inhibitory concentration data was collected from AST for comparisons of accuracy to the ARG markers. The isolates were sub-cultured, and bacterial DNA was extracted for the construction of a whole-genome sequence library. Sequencing was conducted using the Illumina MiSeq v3 kit with 2 x 300 cycles. The resulting reads were trimmed using Trimmomatic v.0.39 and quality checked reads were assembled using SPAdes v.3.13.0 to generate assemblies. An additional 26 Mannheimia haemolytica genomes were downloaded from the Owen et al., 2017 study and 6 samples were obtained from BioProject PRJNA824533. Subsequently, contigs were annotated using RAST v.2.0, and known ARGs were identified by aligning contig sequences against known sequences in the Comprehensive Antibiotic Resistance Database (CARD). The contigs were then partitioned into overlapping segments, each 31 bases in length. These segments, along with phenotypic antimicrobial susceptibility testing (AST) data, were utilized as input for the machine learning algorithm.