Leveraging of Machine Learning to Evaluate Genotypic-Phenotypic Concordance of Pasteurella Multocida Isolated from Bovine Respiratory Disease Cases

Pasteurella multocida is a frequently isolated respiratory pathogen from cattle suffering from bovine respiratory disease (BRD), which is the leading cause of mortality and morbidity on modern-day cattle farms. Treatment involves the use of antimicrobials which have been shown to fail for about 30% of BRD cases, leading to the suspicion that etiologic agents, such as P. multocida, may be resistant. Phenotypic resistance can be confirmed via laboratory antibiotic susceptibility testing (AST) but this requires several days to complete; whereas genotypic resistance could be quickly assessed via nucleic acid assays based on the presence of known antibiotic resistance genes (ARGs). However, ARGs associated with antibiotics to treat BRD, have showed low phenotype-genotype concordance (Owen et al., 2017 https://doi.org/10.1534/g3.117.1137) and therefore, the purpose of this study is to improve P. multocida phenotype-genotype concordance by applying a machine learning algorithm to identify novel genomic sequences that have greater accuracy in predicting resistance to antibiotics commonly used to treat BRD compared to known ARGs. Bacterial samples (n = 104) were obtained from animal disease diagnostic labs who isolated P. multocida from lung tissue or nasal swabs of cattle that presented clinical signs of BRD. AST was performed, and minimum inhibitory concentration was recorded for each isolate. The isolates were then sub cultured and bacterial DNA was extracted and used for the construction of a whole genome sequence library. Sequencing was performed using the Illumina, MiSeq v3 kit, 2 x 600 cycle and output reads were trimmed (Trimmomatic, v.0.39) and assembled (SPAdes, v.3.13.0) to form assemblies that contained contigs that were greater than or equal to 500 bp in length. Additionally, 25 P. multocida sequences from Owen et al., 2017 were downloaded and similarly trimmed and assembled, resulting in 129 bacterial genomes. Two genomes (PM062 and PM073) were suspected to not be P. multocida and consequently, these samples were not used to create resistance prediction models. Contigs were then annotated (RAST, v.2.0) and ARGs were identified by comparing contig sequences to known ARGs sequences in the comprehensive antibiotic resistance database (CARD). Contigs were then split into overlapping 31-base long segments which were used with phenotypic AST data as input data for the ML algorithm.