Antimicrobial resistance (AMR) is a threat to global health, food security, and economic productivity. Infections caused by drug resistant Gram-negative pathogens, such as Escherichia coli, Pseudomonas aeruginosa, and Neisseria gonorrhoeae, are continuously becoming harder to treat due to limited treatment options and long turnaround times for culture-based phenotypic diagnosis. Alternatively, genotypic approaches that exploit whole genome sequencing have the potential to be faster and more accurate. Genotypic approaches rely on using bacterial genomes to predict AMR phenotypes.
I generated a rules-based algorithm and machine learning models using known resistance determinants from bacterial genomes to predict resistance or susceptibility.
I showed that machine learning was superior to a rules-based algorithm and achieved an average accuracy of 94% and 89% for E. coli and P. aeruginosa, respectively. These machine learning models identified novel AMR genotype-phenotype relationships between known resistance determinants and resistance phenotypes, which were experimentally validated.
To identify the parameters that can improve machine learning models, I tested a variety of genetic features, algorithms, and evaluation metrics. I observed an intricate dependency between parameters for AMR prediction performance, illustrating that careful selection of parameters is required to generate accurate AMR prediction models.
A limitation of this work was its prediction of resistance and susceptibility categories, as these are interpretations of minimum inhibitory concentrations defined by clinical breakpoint guidelines. Since multiple guidelines exist, these prediction models are not generalizable, so prediction of MIC values was explored. The average accuracy of my MIC prediction models was 86%, 41%, and 98% for E. coli, P. aeruginosa, and N. gonorrhoea, respectively.
Despite the multifactorial and intricate nature of the resistome, I was able to accurately predict AMR phenotypes for many antibiotics for these pathogens. This is a step towards advanced diagnostic microbiology methods driven by genomics. / Thesis / Doctor of Philosophy (PhD) / Many surgeries, chemotherapy, and transplantation will be impossible if antibiotic resistance is not addressed. Antibiotic misuse, overuse, and time to definitive therapy exacerbate this global health problem. Phenotypic testing determines definitive therapy, but bacterial culturing is slow. A potentially faster and more accurate approach relies on sequencing the pathogen’s genome.
I used machine learning to generate antibiotic resistance prediction models that achieved average accuracies of 94% and 89% for Escherichia coli and Pseudomonas aeruginosa, respectively. These models identified novel relationships between known resistance genes and resistance phenotypes, which were experimentally validated.
Resistance and susceptibility are interpretations of a minimum inhibitory concentration (MIC) using a clinical breakpoint guideline. Since there are different guidelines, I generated MIC prediction models with average accuracies of 86%, 41%, and 98% for E. coli, P. aeruginosa, and Neisseria gonorrhoea, respectively.
My findings work towards a world where clinical sequencing and genomics-based diagnostics are the gold standard.
| Identifer | oai:union.ndltd.org:mcmaster.ca/oai:macsphere.mcmaster.ca:11375/26789 |
| Date | January 2021 |
| Creators | Tsang, Kara K. |
| Contributors | McArthur, Andrew G., Biochemistry and Biomedical Sciences |
| Source Sets | McMaster University |
| Language | English |
| Detected Language | English |
| Type | Thesis |
Page generated in 0.0017 seconds