Clinical Implications of Genomic Variation in Mycobacterium tuberculosis: Association, Prediction, and Within-Host Diversity

Abstract

Tuberculosis (TB), caused by the bacterium Mycobacterium tuberculosis, is the leading infectious disease killer globally. Whole genome sequencing (WGS) has transformed TB surveillance, resistance diagnosis, and outbreak detection and control. The cost of sequencing continues to fall, but WGS is still expensive for routine clinical care in high-burden TB settings. In this work, we show that quantitative metrics derived from WGS at treatment onset are associated with long-term clinical TB outcomes. This suggests their potential utility as baseline indicators to inform risk stratification and clinical management early in the course of treatment. The World Health Organization (WHO) has endorsed genotypic identification of antibiotic resistance in M. tuberculosis by detecting mutations known to confer resistance. Developing these diagnostics requires a high-confidence catalog of mutations. In chapter 1, we apply multivariate regression models to build a catalog of resistance-associated mutations and benchmark it against the 2023 version of the WHO-endorsed catalog. We find that regression models increase sensitivity for rare variants and accurately learn the individual effects of co-occurring epistatic mutations. In chapters 2 and 3, we study the association of WGS-derived metrics with clinical TB outcomes. We use data from a cohort of 452 individuals in South Africa with predominantly drug-susceptible TB. In chapter 2, we build deep learning models to predict quantitative antibiotic resistance (minimum inhibitory concentrations, MICs) for 8 drugs and show that interpretable deep learning models can be used to generate hypotheses about less well-characterized proteins. In chapter 3, we use short- and long-read hybrid assemblies to develop a pipeline to accurately identify unfixed single nucleotide variants (SNVs) and indels. We show that at baseline, both predicted rifampicin MIC below the resistance breakpoint and total number of unfixed SNVs are associated with worse patient outcomes. We also find that unfixed SNVs and indels are concentrated in transcription factors, transporter proteins, and two component regulatory systems. These findings suggest that genomic variation below the threshold of drug resistance contains information relevant to disease progression and treatment response. More broadly, this work illustrates how quantitative analysis of low-frequency variation can extend the clinical utility of WGS beyond categorical resistance prediction.