AMRscope

AMRscope predicts the likelihood that a single mutation in a bacterial protein confers antimicrobial resistance (AMR), turning protein-language-model representations of point mutations into a risk score for surveillance. Antimicrobial resistance is among the most pressing global health threats, and it can arise through diverse mechanisms, including genomic mutations that stop antibiotics from reaching or acting on their targets. To contain its spread, surveillance systems must flag not only known resistance markers but emerging ones — the mutations no database has yet catalogued.

The model was developed by Jennifer J. Wood and Nicholas Furnham at the London School of Hygiene & Tropical Medicine, with Stephanie Portelli and David B. Ascher at the University of Queensland, and posted to bioRxiv in September 2025. AMRscope is a classifier trained on ESM-2 protein language model embeddings of single mutations, paired with a rigorous evaluation framework. It is a trained instance of the more general MUTscope pipeline for predicting the effects of single amino acid substitutions.

Critically, beyond scoring held-out mutations within its training distribution, AMRscope is evaluated with splitting strategies that test transfer to entirely unseen organisms and genes — the setting that matters for catching novel resistance.

#Key Features

• PLM-embedding classifier: Represents each single mutation with ESM-2 embeddings and learns a classifier over them, avoiding hand-engineered features while staying light and fast to train and run.
• Cross-organism transfer: Dedicated splitting strategies show the model transfers predictive power to unseen organisms and genes, not merely to random held-out mutations.
• Priority-pathogen coverage: Applied across antibiotic-interacting proteins from multiple species, including WHO priority pathogens such as rifampicin-resistant M. tuberculosis and carbapenem-resistant P. aeruginosa.
• Interpretable outputs: In silico deep mutational scanning and structural mapping recover known resistance-associated regions and highlight new candidate sites, aiding interpretation alongside database matching.

#Technical Details

AMRscope embeds single amino acid substitutions using the ESM-2 protein language model and trains a classifier head to predict resistance likelihood, with the underlying MUTscope framework supporting interchangeable classifier heads (MLP, CNN, and others). On random splits it reaches an accuracy of 0.88, an F1 of 0.87, and a Matthews correlation coefficient of 0.75; additional organism- and gene-based splits demonstrate transfer of predictive power to entities unseen during training. In silico deep mutational scanning across the target proteins, combined with structural mapping, recovers known resistance-associated regions and surfaces new candidate sites. Code for the MUTscope pipeline, of which AMRscope is a trained instance, is released on GitHub under an MIT license.

#Applications

AMRscope is designed for AMR surveillance and proactive response, giving clinicians and public-health agencies an interpretable, scalable risk score that complements database matching and resistance-element detection tools. Because it generalizes to unseen organisms and genes, it can help flag emerging resistance mutations before they are formally catalogued, prioritizing variants for laboratory confirmation and informing stewardship decisions across bacterial pathogens.

#Impact

By demonstrating that a classifier over ESM-2 embeddings can recover known resistance regions and transfer to unseen organisms and genes, AMRscope shows how frozen protein language models can power practical, interpretable AMR surveillance. Its architectural thinness is a strength for deployment, and its risk-based outputs are meant to sit alongside existing detection tools rather than replace them. As a preprint awaiting peer review, released through the open-source MUTscope pipeline, its broader adoption will depend on independent validation and clear distribution of the specific AMRscope checkpoint.