New preprint: validating antibiotic resistance prediction in our Myco pipeline

Over the last 18 months or so we’ve been designing, coding and testing a Mycobacterial genetics processing pipeline (which we call Myco for short). This pipeline has been deployed on the EIT Pathogena cloud platform and is free for academic researchers and LMICs to use. One can upload raw genetics files (FASTQ) from a putative Mycobacterial sample and the pipeline will determine what Mycobacterial species are present and, if there are enough M. tuberculosis reads, produce a consensus genome to which it will apply the second edition of the WHO catalogue of resistance-associated variants (WHOv2) to produce an antibiogram. Finally it will compare the genome to all others uploaded and return a list of other samples that could be epidemiologically related (that the user has permission to view).

Since one cannot directly apply WHOv2 to a sample, one must first translate it into a form that can be used by a computational tool. In this preprint, we focus on assessing the performance of our translation for 15 different antibiotics using a diverse testset of 2,663 publicly available M. tuberculosis samples — these were chosen to have as balanced a resistance profile as possible. We find comparable performance to that reported in the WHOv2 report. We also make several enhancements; these include return a result of Fail if there are insufficient reads at a genetic locus associated with resistance, returning an Unknown result if a mutation in a resistance gene not in the catalogue is encountered and finally calling Resistant if three or more reads at a locus support the identification of one the resistance-associated variants from the catalogues. The last improvement is based on work done by Alice Brankin during her DPhil.

Finally, since the dataset of 2,663 samples can be downloaded by anyone and our results can also be found in the attendant GitHub repository, we hope that this could form the basis of testing and benchmarking new catalogues and tools. This is not currently possible since the WHOv2 Training Dataset is not publicly available.

You can read the preprint here.