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

Tessa Ruth Sheets, Purdue University

Abstract

Pasteurella multocida is a respiratory pathogen that is frequently isolated from cattle suffering from bovine respiratory disease (BRD), the leading cause of mortality and morbidity on modernday 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. Genotypic resistance could be quickly assessed via nucleic acid assays based on the presence of known antibiotic resistance genes (ARGs). In human medicine, resistant genes associated with common antibiotics (i.e., ampicillin and penicillin) in common pathogens (i.e., Salmonella) are very accurate in predicting phenotypic resistance; however, ARGs associated with antibiotics used to treat BRD, such as enrofloxacin and tulathromycin, have shown low genotype-phenotype concordance. Hence, this study aims to improve P. multocida genotype-phenotype concordance by applying a machine learning (ML) algorithm to identify novel genomic sequences (biomarkers) that have greater accuracy in predicting resistance to antibiotics commonly used to treat BRD compared to known ARGs. Cultures of P. multocida were isolated from cattle with clinical signs of BRD. Antibiotic susceptibility testing was performed and recorded for each isolate. Genomes were sequenced and assembled, followed by annotating and identifying ARGs using the comprehensive antibiotic resistance database (CARD). Assembled genomes were then split into 31-base long segments (31- mers), and these segments along with phenotypic antibiotic susceptibility were used as input data for the ML algorithm. Important genomic biomarkers for four out of the six tested antibiotics were found to have greater accuracy when predicting resistance phenotype compared to known ARGs. The biomarker for enrofloxacin had the highest accuracy of 100% whereas the accuracy for the tulathromycin biomarker was 81% but was still greater than the accuracy given by ARGs of 63%. On the other hand, resistance genes for florfenicol and tetracycline showed greater genotypephenotype concordance, with accuracies of 95% and 91%, respectively. Annotations to important rulesets determined by ML were associated with clustered regularly interspaced short palindromic repeats (CRISPR) sequences, ligases that function to recycle murein into the peptidoglycan (PDG) layer, and transferases that control the synthesis and modulation of the lipopolysaccharide (LPS). External validation revealed that phenotypic resistance could be accurately predicted for danofloxacin and enrofloxacin using genomic biomarkers determined by ML, and for florfenicol using the floR gene. This study demonstrated that genomic biomarkers determined by ML can provide an accurate prediction of antibiotic resistance within Pasteurella multocidaisolates. Assays could be developed to target ML-generated biomarkers and known ARGs to predict resistance in sick animals and to limit treatment failures associated with antibiotic resistance in cattle suffering from BRD.

Degree

M.Sc.

Advisors

Johnson, Purdue University.

Subject Area

Pharmacology|Virology

COinS