Background: Clinical bedaquiline resistance predominantly involves mutations in mmpR5 (Rv0678). However, mmpR5 resistance-associated variants (RAVs) have a variable relationship with phenotypic M. tuberculosis resistance. We performed a systematic review to (1) assess the maximal sensitivity of sequencing bedaquiline resistance-associated genes and (2) evaluate the association between RAVs and phenotypic resistance, using traditional and machine-based learning techniques.
Methods: We screened public databases for articles published until October 2022. Eligible studies performed sequencing of at least mmpR5 and atpE on clinically-sourced M. tuberculosis isolates and measured bedaquiline minimum inhibitory concentrations (MICs). We performed genetic analysis for identification of phenotypic resistance and determined the association of RAVs with resistance. Machine-based learning methods were employed to define test characteristics of optimised sets of RAVs, and mmpR5 mutations were mapped to the protein structure to highlight mechanisms of resistance.
Results: Eighteen eligible studies were identified, comprising 975 M. tuberculosis isolates containing ≥1 potential RAV (mutation in mmpR5, atpE, atpB or pepQ), with 201 (20.6%) demonstrating phenotypic bedaquiline resistance. 84/285 (29.5%) resistant isolates had no candidate gene mutation. Sensitivity and positive predictive value of taking an 'any mutation' approach was 69% and 14% respectively. Thirteen mutations, all in mmpR5, had a significant association with a resistant MIC (adjusted p<0.05). Gradient-boosted machine classifier models for predicting intermediate/resistant and resistant phenotypes both had receiver operator characteristic c-statistics of 0.73. Frameshift mutations clustered in the alpha 1 helix DNA binding domain, and substitutions in the alpha 2 and 3 helix hinge region and in the alpha 4 helix binding domain.
Discussion: Sequencing candidate genes is insufficiently sensitive to diagnose clinical bedaquiline resistance, but where identified a limited number of mutations should be assumed to be associated with resistance. Genomic tools are most likely to be effective in combination with rapid phenotypic diagnostics.