Protein-structure AI predicts pyrazinamide resistance

Tuberculosis remains a global health challenge and pyrazinamide is an important first-line antibiotic used to treat it. Resistance to pyrazinamide is driven mainly by changes in the bacterial pncA gene, but not all pncA mutations have the same effect, and predicting which ones cause resistance is difficult. Traditional machine learning approaches have had some success, yet they struggle to include detailed 3-dimensional protein information that can explain how a mutation alters drug response. In work led by Philip W. Fowler, researchers explored a different approach: a graph convolutional network that treats each PncA protein variant as a network of amino acids linked by their spatial relationships. By turning protein structures into graphs and adding biochemical information about individual residues, the team aimed to directly capture how changes in PncA shape and chemistry influence pyrazinamide resistance. This study was presented as a proof-of-concept to show that combining structural detail and residue-level features in a single model can improve predictions of whether a given mutation will make Mycobacterium tuberculosis resistant to pyrazinamide.

To build their predictor, the team trained a graph convolutional network (GCN) on PncA variants that contained missense mutations. Each variant was converted into an amino acid-level graph: edges were calculated from 3-dimensional spatial proximity and node features came from chemical properties and mutation meta-predictors. The researchers used AlphaFold2 to generate predicted structures of the PncA variants and used those structures to make the protein graphs. When they compared predicted structures, resistant PncA variants showed greater deviation from the wild-type structure than susceptible variants. On a held-out test set the model produced an F1 score of 81.6 %, sensitivity of 81.6 % and specificity of 80.4 %. The GCN either matched or exceeded the performance of a published set of traditional machine learning models. The authors also showed that both the structural graph connectivity and the node features contributed significantly to performance, and additional train/test splits demonstrated the GCN’s ability to generalise to mutations in unseen positions and structural regions, despite being trained on a small dataset with little variation.

The study highlights how graph-based deep learning can use protein structure and biochemical features to make accurate predictions about antimicrobial resistance. By relying on AlphaFold2-predicted structures and residue-level features, the approach offers a way to move beyond sequence-only methods and to reason about how specific mutations change the shape and chemistry of PncA in ways that affect pyrazinamide activity. Because the work was presented as a proof-of-concept, it does not claim immediate clinical readiness, but it does suggest a promising direction: applying similar models to more genetically diverse pathogens could help predict more complex patterns of antimicrobial resistance. The findings point to a future where structural models and graph neural networks augment laboratory testing and traditional genetic analyses, offering faster, more interpretable predictions that can guide research into resistance mechanisms and potentially improve decision-making around antibiotic treatment strategies.