Investigating the genomic basis of antimicrobial resistance in Mycobacterium tuberculosis (Mtb) using genome-wide methodologies

Characterizing the drug resistance mutations that have evolved in Mycobacterium tuberculosis (Mtb), has important implications for control of tuberculosis (TB) disease, through more accurate and timely use of therapy. Whole genome sequencing of Mtb can assist this characterization by providing insights into loci and specific mutations underlying drug resistance and the transmission success that enables their spread. We hypothesised that genetic variation outside of known resistance-conferring mutations might give additional information concerning drug resistance and fitness. Firstly, we explored the effect of lineage on the identification of drug resistance associations, applying novel lineage level genome-wide association study (GWAS) and convergence-based (PhyC) methods to drug resistance phenotypes of a global dataset of Mtb lineages 2 and 4. We identified known drug resistance variants and novel associations, uniquely identifying associations for lineage-specific GWAS analyses and reporting 17 novel associations between antimicrobial resistance phenotypes and Mtb genomic variants, demonstrating the utility of lineage-specific GWAS. To further examine the genomic basis of extensively drug resistant (XDR)-TB, we next applied the GWAS and PhyC techniques to a global dataset of 18,255 Mtb isolates. Through GWAS we identified 20 loci in novel associations within highly drug-resistant Mtb strains. Cluster-based GWAS and a lack of overlap with associations identified through convergent-evolution-based analyses confirmed that many such associations have been driven by transmission in outbreaks of XDR-TB. We then investigated the feasibility of applying a learning classifier system to this dataset to predict rifampicin resistance and discover candidate loci for novel involvement, finally enabling a sensitivity of 93.7% and a specificity of 94.8% of rifampicin resistance prediction. Finally, we applied this methodology to the XDR phenotype in lineages 2 and 4 of a global dataset (n=13,270), achieving high accuracy of prediction and identifying a number of candidate loci for involvement in XDR, including candidates for epistasis.