Automated catalogue tracks TB drug resistance variants

Tuberculosis remains a major global health threat, and predicting which drug will work against a given bacterial strain increasingly relies on reading the bacterium’s DNA. Catalogues that link genetic changes to drug resistance are the foundation for whole-genome sequencing (WGS)-based predictions, molecular diagnostics and public health surveillance. But the current gold standard resources, including the catalogues released by WHO, have limitations: underlying data are not fully released and the final lists can be hard to interpret or reproduce. To address this, Philip W. Fowler and colleagues developed an automated, transparent approach called catomatic and used it to build catomatic-1. The authors started from the very same 39,358 samples used to construct the first edition of the WHO catalogue (WHOv1) and applied a clear statistical rule - a binomial test - to associate individual genetic variants with resistance or susceptibility. They performed a sensitivity analysis to tune parameters and then benchmarked catomatic-1 on an independent Validation Dataset of 14,380 isolates, showing a practical, reproducible path to update and extend resistance catalogues.

catomatic is an automated pipeline that uses a binomial test to classify variants as associated with resistance or susceptibility. Trained on the WHOv1 sample set and benchmarked on the independent Validation Dataset, catomatic-1 algorithmically classified 1,329 genetic variants overall, with counts ranging from five for linezolid to 440 for pyrazinamide. The team explored statistical and bioinformatic parameters per drug and found that adjusting read-support thresholds — for example, detecting resistant variants with low read support — improved sensitivity for all drugs. Importantly, WHOv1 included additional generalisable rules added by a panel of experts to increase predictive coverage, but those expert rules reduce reproducibility. Despite using simpler statistics and no expert-rule augmentation, catomatic-1 achieved comparable sensitivity, specificity and definitive prediction rates to WHOv1, with sensitivities for first-line agents above 88% on the Validation Dataset. The tool produces machine-readable outputs (CSV/JSON) and optimises parameters per drug, making it practical for diagnostics and surveillance use.

The work shows that performant resistance catalogues for Mycobacterium tuberculosis can be built rapidly using transparent, reproducible statistical methods. Because the process is automated, catalogue content can evolve as new data arrive, but that also means careful versioning, machine- and human-readable formats, and open access are essential so users know which data and thresholds underlie each release. The authors argue policymakers and guideline developers must weigh the benefits of adding expert rules against the loss of reproducibility those rules can cause. For laboratories, diagnostic test manufacturers, researchers and public health programmes, a reproducible pipeline like catomatic offers the choice to select the statistical support level that suits their use-case. Future work outlined by the team includes expanding datasets, integrating minimum inhibitory concentration (MIC) data, and establishing consensus workflows so routine catalogue updates remain transparent and dependable. Funding came from the National Institute for Health and Care Research (NIHR), Engineering and Physics Sciences Research Council (EPSRC) and ORACLE Corporation.