Summary

This work presents, to our knowledge, the first application of structural deep learning - the explicit modelling of protein structure using a graph convolutional network (GCN) - to antimicrobial resistance prediction, demonstrated in M. tuberculosis. By representing PncA variants as residue-level structural graphs derived from AlphaFold2 predictions, the model can leverage both structural context and biochemical features to predict pyrazinamide resistance.

We demonstrate that the model either matches or outperforms the best previously published machine learning approaches, notably in the most difficult task of out-of-distribution resistance prediction of mutations in unseen structural regions, where it shows a strong ability to generalise.

This study shows how structural deep learning can move AMR prediction beyond catalogues and towards inference for novel mutations. This proof of concept establishes a framework that can now be extended to more genetically diverse pathogens, where structural context may offer even greater gains.