Predicting hidden antibiotic resistance with AI-trained protein models

Antimicrobial resistance (AMR) is described as one of the most pressing global health threats of the 21st century, with the potential to undermine modern medicine. Resistance can arise through diverse mechanisms, including genomic mutations that prevent antibiotics from reaching or acting on their targets, so surveillance must detect both known and emerging resistance markers. To address that challenge, Nicholas Furnham and colleagues developed AMRscope, a predictive tool built around protein language model information. AMRscope is trained on ESM2 protein language model embeddings of single mutations and paired with a rigorous evaluation framework to estimate the likelihood that a given mutation will confer resistance. The team applied the tool across antibiotic-interacting proteins from different bacterial species, including WHO priority pathogens. The study highlights examples such as rifampicin-resistant M. tuberculosis and carbapenem-resistant P. Aeruginosa to show the approach works on clinically important targets. By focusing on single mutation effects in protein targets, AMRscope aims to extend surveillance beyond exact database matches and help identify mutations that may signal emerging resistance.

The core method uses embeddings from the ESM2 protein language model to represent single amino-acid changes, and trains a model to predict resistance likelihood from those representations. The work pairs this modeling with a careful evaluation framework: on random data splits the method achieves accuracy, F1 and MCC of 0.88, 0.87 and 0.75, respectively, indicating strong performance on held-out examples. The authors also applied additional splitting strategies to test generalization, and found that predictive power transfers to unseen organisms or genes, suggesting the model can work beyond the exact examples it was trained on. The study further uses in silico deep mutational scanning and structural mapping across targets to recover known resistance-associated regions and to highlight new candidate sites. AMRscope produces risk-based outputs designed to complement database matching and resistance element detection tools, offering interpretable scores that can be scaled across proteins and species.

The implications of this approach center on improved surveillance and a more proactive response to AMR. By estimating the resistance likelihood of single mutations across antibiotic-interacting proteins, AMRscope can surface variants that are not yet catalogued in existing databases, helping clinicians and public health agencies prioritize follow-up and monitoring. The transfer of predictive power to unseen organisms or genes suggests the method could be applied to a wide range of bacterial targets, including WHO priority pathogens like rifampicin-resistant M. tuberculosis and carbapenem-resistant P. Aeruginosa. Structural mapping and deep mutational scanning help make the predictions interpretable, pointing to regions of proteins that warrant laboratory validation. Overall, the tool is presented as an interpretable, scalable complement to existing surveillance workflows, intended to aid detection of both known and emergent resistance markers and support more informed public health action.