New model predicts bedaquiline resistance linked to Rv0678 variants

Bedaquiline resistance is emerging around the world and threatens the usefulness of the novel short all-oral regimens for rifampicin-resistant tuberculosis. To confront that problem, a team led by Annelies Van Rie reviewed the scientific literature and collected data on genetic changes tied to reduced susceptibility. They focused on the Rv0678 gene, which when altered can affect the bacterium Mycobacterium tuberculosis and change how it responds to bedaquiline. From reported cases they identified 62 Rv0678 missense variants found in 136 Mycobacterium tuberculosis isolates and measured a set of features that describe each change. Rather than relying on single observations, the researchers took a feature-based approach, quantifying 13 sequence, biochemical, and structural characteristics for each variant. Their goal was to build a practical, computational — in other words, in-silico — tool that predicts whether a particular Rv0678 variant is likely to produce a bedaquiline-resistant phenotype. By combining careful data collection with modern computational analysis, the study aims to give clinicians and laboratories better information about which genetic changes matter for treatment decisions.

The team used rigorous machine learning methods to find which of the 13 features best distinguish variants that alter bedaquiline susceptibility. Among the characteristics they measured were evolutionary conservation scores that indicate how preserved an amino acid is across species, and structural measurements such as the shortest atomic distance from the altered residue to key functional sites in the protein. These features reflect both the likely biological importance of a change and how it might disrupt protein function. After testing different combinations, the researchers produced a final 5-feature model that performed well: it achieved an ROC-AUC of 0.826 and classified the bedaquiline phenotype with sensitivity 87.1% (95% CI, 78.3–92.6) and specificity 88.2% (95% CI, 76.6–94.5). They also examined how missense variants map onto the mmpR5 protein structure and function to understand likely mechanisms. Performance was lower when the model was tested on external datasets, a shortcoming the authors attribute to measurement error introduced by using diverse phenotypic methods across studies.

The findings point to a practical way to improve interpretation of genetic test results for rifampicin-resistant tuberculosis. Integrating the five-feature in-silico model into variant interpretation software could help laboratories predict the effect of Rv0678 variants more consistently than subjective judgment alone. That would make it easier to flag cases where bedaquiline may be less effective and could guide treatment choices in settings facing growing resistance. The study also highlights important caveats: differences in how laboratories measure phenotypic resistance can reduce model performance in external validation, so harmonized testing and further validation are needed. In short, this work suggests a clear next step — incorporate these predictive features into decision tools and continue testing them on broader datasets — to help clinicians and public health teams respond faster to emerging bedaquiline resistance.