Language-guided domain adaptation improves protein sequence rep-resentations for antimicrobial resistance analysis

Abstract

Antimicrobial resistance (AMR), particularly antibiotic resistance in bacterial pathogens, complicates infectious disease management. Because bacterial resistance is often mediated by protein functions, including antibiotic inactivation and efflux, protein sequence analysis provides a molecular basis for digital AMR surveillance and therapeutic intervention. We developed BacLLM, a language guided protein representation framework that integrates annotation-derived functional knowledge into protein language model representations for AMR analysis in WHO priority bacterial pathogens. Through sequence enhancement, semantic enrichment of UniProt annotations, and cross modal contrastive learning, BacLLM transfers LLM encoded annotation priors into the sequence encoder. Compared with baselines, BacLLM provided more informative representations for resistance-gene identification and mechanism prediction, captured resistance associated mutational signals at single-residue resolution, and preserved taxonomic relationships not reflected by phenotype based clinical groupings. Moreover, BacLLM recovered resistance associated functional organization within a TolC-centered protein association network, prioritizing proteins with low sequence identity to TolC but predicted structural similarity to TolC. Extending from AMR characterization to therapeutic prioritization, BacLLM also improved strain-level phage host matching for phage therapy candidate prioritization. Together, these findings establish BacLLM as a protein representation framework supporting AMR surveillance, functional characterization, and therapeutic candidate prioritization.