Philip Fowler

I am a group leader in the Modernising Medical Microbiology Unit which is based in the John Radcliffe Hospital in Oxford. Our goal is to develop computational methods that can predict antibiotic resistance. 

Most often this is in the context of genetics-based clinical microbiology. The central idea here is to infer which antibiotics can be used to treat an infection by examining the genetics of a clinical sample. Our exemplar pathogen is Mycobacterium tuberculosis complex (MTBC) which is the aetiology agent of tuberculosis due to its very slow growth rate and relatively straightforward genetics, although we are now consider all Mycobacterial species thanks to our HPRU which makes it straightforward to collaborate with UKHSA. Our Mycobacterial genetics pipeline is freely-available and can process both short- and long-reads and yields species information, the genome, to which antibiotics the organisms is likely to be susceptible and a list of other samples that are likely to be epidemiologically related. The pipeline is deployed on a commercial cloud platform that we call GPAS which simplifies the whole process for end users and allows us to process thousands of samples per day.

We also curate data to inform the development of our prediction approaches; starting from the earlier CRyPTIC project, I have amassed to date 53,897 MTB samples each of which has a genome and a range of different phenotypic antibiotic susceptibility data (AST). Version 1.1.1 of  this dataset was used by the WHO to create their catalogue of resistance associated mutations in MTBC which was first released in June 2021 and updated in November 2023. As part of the CRyPTIC project I developed image processing software called AMyGDA that determined the minimum inhibitory concentrations of the dozen drugs on a 96-well broth micro dilution plate; this also enabled a very large and successful citizen science project, BashTheBug, through which 46,427 volunteers classified 4,746,420 images of MTB growing on different antibiotics. More recently we have retrained some off-the-shelf deep learning models to classify images of the 96-well plates.

The state-of-the-art approach associates genetic variants with the phenotypic drug susceptibility outcome and is therefore inferential and therefore cannot make any prediction for variants not in its training dataset. We are therefore developing predictive methods. These fall into two main classes: (1) protein structure-based machine learning and (2) calculating the effect of the mutation on the binding free energy of the antibiotic using molecular simulations. The former covers a range of machine learning methods, from relatively straightforward methods like random forest through to more advanced methods like graph-based convolution neural networks (gCNNs). Our hope with the latter is to begin to move into other species (like the ESKAPE pathogens) because we are predicting based off sequence rather than mutation. 

If you'd like to know more please see our group website or get in touch.