Predicting fluoroquinolone resistance in Mycobacterium tuberculosis

Thesis

Tuberculosis killed an estimated 1.5 million people in 2020 and the spread of drug-resistant strains is an increasing threat. Fluoroquinolones are among the safest and most effective drugs used to treat drug-resistant tuberculosis, but fluoroquinolone resistance has emerged. Rapid resistance diagnosis is key to provide patients with effective treatment and monitor the spread of resistant strains. Sequence-based tools, including whole genome sequencing (WGS), provide a promising solution whereby known resistance associated mutations can be quickly detected and used to infer resistance. Several catalogues of mutations with known associations have now been produced, however, these catalogues will never be exhaustive owing to the inevitable emergence of novel and rare resistant mutations.

In this thesis I explore a geographically diverse dataset of thousands of Mycobacterium tuberculosis isolates with WGS and phenotypic resistance measurements, to improve our understanding of patterns associated with fluoroquinolone resistance. I investigate the reliability of the assumptions of current sequence-based diagnostics and show how resistance patterns can affect the performance of predictive tools. I apply and evaluate two methods, that utilise the chemical and structural interactions resulting from genetic mutations, to predict the effects of mutations on fluoroquinolone resistance; machine learning algorithms and free energy calculation.

I find that fluoroquinolone resistance is widespread and the associations with resistance are complex; genetic and geographic background, resistance to other antitubercular drugs and minor populations with resistance associated mutations are important. I show that structure-based predictive methods, and the machine learning approach particularly, can successfully predict fluoroquinolone resistance and susceptibility. Such tools could increase the success of resistance prediction from WGS data by complimenting the catalogue-based predictive approach and predicting the effects of novel mutations. Overall, this work shows that genetics-based predictive diagnostics have the potential to provide personalised, effective treatment regimens for tuberculosis.

Authors: Brankin, AE

• DOI: 10.5287/ora-qmvn44gy6
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: free energy calculations; fluoroquinolones; antibiotic resistance; antimicrobial resistance; tuberculosis; machine learning
• Subjects: Diagnostic microbiology; Supervised learning (Machine learning); Clinical medicine--Research; Research; Bioinformatics; Statistics