New preprint: rapidly and reproducibly building resistant catalogues for M. tuberculosis

The CRyPTIC project carried out many exciting research projects but it never quite got around to building a catalogue of resistance- and susceptible-associated genetic variants in M. tuberculosis, in part because near the end of the project we donated our entire dataset to the Seq&Treat project who used it to build the first edition of the WHO catalogue of mutations in M. tuberculosis.

Both editions of the catalogues published by the WHO, whilst representing the current gold standard, use a mixture of statistical association testing and the addition of Expert Rules. This, in combination with the fact that the most recent dataset is not publicly available, makes them difficult to reproduce. In addition the catalogue is at least two separate artefacts; a very large Excel spreadsheet contains the results of the statistical association testing and some Expert Rules but the report (which is a PDF) contains some additional rules that must also be taken into account. Together these factors make it very challenging to parse, understand and apply these catalogues.

Elsewhere we describe our efforts to translate the second edition of the WHO catalogue into our GARC grammar which ensures it is both human- and computer-readable and confines the catalogue to a single CSV file and then check our process by assessing its performance and comparing it to TB-Profiler another well known tool. This catalogue can then be ingested by a software tool we’ve written gnomonicus.

Recently we’ve made the CRyPTIC project datasets available via Zenodo as the original FTP site only contained the original version of the dataset and we wanted to make it easier to get all versions.

Finally, Dylan Adlard has written a Python package, catomatic, that, if given data tables in the format produced by CRyPTIC, applies statistical association testing to classify genetic variants with either resistance or susceptibility to a specific antibiotic.

This preprint brings all this together and allows us to create Training and Test datasets using v3.4.0 of the CRyPTIC dataset – the Training dataset is purposefully chosen to be identical to that used to build the first version of the WHO catalogue so we can fairly compare the performance of both catalogues on the independent Validation dataset.

So it has taken a while and builds on a lot of work making the datasets publicly available and software engineering, but we can now rapidly and reproducibly build resistance catalogues in a format that is both human- and computer-readable. This opens up a few interesting possibilities, for example building lineage-specific catalogues or catalogues built using samples from one geographic region. Or even building the first catalogue of resistance mutations for common NTM species.