Closing the Gaps - Improving genetics-based predictions for antimicrobial resistance in Mycobacterium tuberculosis and Escherichia coli

<em>“If we use antibiotics when not needed, we may not have them when they are most needed.” </em>— Tom Frieden, former director of the U.S. Centers for Disease Control and Prevention. <br><br>This statement remains acutely relevant as antimicrobial resistance (AMR) continues to rise globally. Inappropriate or delayed antibiotic use accelerates the spread of resistance, underscoring that effective diagnostics are as essential as the development of new antimicrobial agents. The central aim of this thesis is improving the prediction of antibiotic resistance, with particular emphasis on whole genome sequencing-based antimicrobial susceptibility testing (WGS-AST). <br><br>WGS-AST has transformed clinical microbiology by enabling rapid and comprehensive resistance prediction relative to traditional phenotypic testing, particularly for M. tuberculosis. However, discrepancies between phenotypic and genotypic results persist, limiting the clinical reliability of genomic prediction. This work investigates the sources of these discrepancies across three main areas: (i) the role of compensatory and fitness-related mutations in shaping resistance and their potential predictive value, (ii) the detection of resistant subpopulations and within-host diversity in both M. tuberculosis and Enterobacteriaceae, and (iii) the application of machine learning approaches, including graph-based models, to predict resistance from sequence, structural, and physicochemical data. <br><br>By combining evolutionary insights, quantitative analyses of within-sample diversity, and predictive modelling, this thesis outlines both the potential and the current limitations of WGS-AST. The findings demonstrate that compensatory mutations can serve as highly specific indicators of resistance, that resistant subpopulations have important clinical and epidemiological implications, and that machine learning models show promise but remain constrained by the underlying genetic architecture and available training data. Together, these results contribute to closing the gap between phenotypic and genotypic testing and advance the development of diagnostic frameworks that are biologically informed and clinically actionable.