tsAMP: a strain-level antimicrobial peptide identification framework based on large language models and pathogen genomic variation

Introduction Facing the global threat of multidrug-resistant bacteria, antimicrobial peptides (AMPs) represent a promising alternative to conventional antibiotics. Methods To improve computational AMP identification and accuracy of strain-level MIC prediction, we developed tsAMP, a comprehensive framework integrating the ESM-1v protein language model with multidimensional feature extraction. The model was trained on AMP and metagenome-derived non-AMP sequences. Results tsAMP achieved an F1-score of 0.958 for AMP identification, outperforming state-of-the-art tools. For bacterial inhibition prediction, tsAMP consistently maintained F1-scores above 0.8 across 33 pathogenic species. In strain-specific MIC prediction, it attained high performance (MSE = 0.214, R 2 = 0.634) for 10 bacterial species’ strains. To assess predictive reliability, the model was benchmarked against published experimentally determined MIC values for AMPs targeting Micrococcus luteus , yielding low prediction error (MSE = 0.1489) and strong ranking consistency (NDCG = 0.791). Computational benchmarking against published relative MIC data for diverse E. coli strains further demonstrated the model’s ranking accuracy (NDCG > 0.85) and consistent strain-level differentiation. Applied to the Mgnify_genome database, tsAMP identified 8,277 putative AMP candidates in silico and revealed distinct predicted antimicrobial activity patterns across pathogens. Discussion tsAMP provides a computational framework to facilitate the identification of AMP candidates and support prioritization for downstream experimental characterization. The code is available on GitHub at https://github.com/YangLab-BUPT/tsAMP .