New paper: Quantitative drug susceptibility testing for M. tuberculosis using unassembled sequencing data and machine learning

This is the last paper from the initial set of CRyPTIC publications following the project’s data freeze in April 2020. The consortium takes a difference approach to that of (i) mapping the reads, (ii) look up the genetic mutations in a catalogue and (iii) return the predictions and instead trained a tree-based extreme gradient-boosted machine learning model.

Since the minimum inhibitory concentration (MIC) was the label, the appropriate metrics are exact and essential agreement which mean “get the same MIC” and “get within one doubling dilution of the MIC”. The essential agreement is good for some drugs like ethambutol which have moderate sensitivities using the traditional binary approach which is expected due to their MIC distribution being almost unimodal. Also the good performance of the fluoroquinolones suggests that the model is able, in part at least, to learn the presence of minor alleles / subpopulations which we have shown elsewhere to be important for this class of drugs.

But seriously: one figure and three tables? All those numbers in tables aren’t exactly easy to read and what I do I put for the thumbnail? (PWF can say this as technically he is an author and therefore it is partly his responsibility and therefore fault).