P048

The inherent resistance of Mycobacterium tuberculosis to β-lactams has meant that the use of these antibiotics to treat tuberculosis has historically not been of clinical significance. Recently BlaC, the chromosomally encoded β-lactamase of M. tuberculosis and the primary cause of this resistance, has been recognised as a potential therapeutic target for β-lactamase inhibitors such as clavulanic acid, thereby enabling the action of β-lactams. However, there is concern that resistance to β-lactamase inhibitors may rapidly arise. We have therefore investigated the natural genetic variation present in blaC in >75,000 genomes using the CRyPTIC dataset, a data compendium of M. tuberculosis samples with mutation and phylogenetic lineage information, and report our findings.

Machine learning models have the potential to predict antimicrobial resistance but require sufficiently large datasets with enough variation – we have therefore also investigated what genetic variation is required for a graph-based neural network (GNN) to begin to learn resistance in blaC. This was done by synthetically generating datasets of M. tuberculosis BlaC sequences with differing characteristics. Resistant mutations in the datasets were determined from previous in vitro studies that identified amino acid substitutions in M. tuberculosis BlaC that conferred resistance to clavulanic acid. The GNN used chemistry-based amino acid features and derived its network connectivity from the protein structures of BlaC samples, so that it could infer awareness of the structural and chemical changes arising from the mutations.