ESM2AMP: An ESM-2-Based Deep Learning Framework for Antimicrobial Peptide Identification and MIC Prediction from Bohai Sea Metagenomes

Antimicrobial peptides (AMPs) are key effector molecules of innate immunity and promising alternatives to conventional antibiotics. However, the discovery and functional characterization of novel AMPs remain constrained by the limitations of traditional experimental approaches. In this study, we present ESM2AMP, a deep learning framework leveraging the ESM-2 protein language model for high-accuracy AMP identification and quantitative activity prediction. The framework consists of two modules: ESM2AMP-I, a stacking ensemble classifier integrating MLP, XGBoost, Random Forest, and SVM trained on 2560-dimensional ESM-2 embeddings, which achieved an accuracy of 0.975 and an F1-score of 0.968 on an independent test set, substantially outperforming existing methods including Diff-AMP, iAMPCN, and AMPfinder; and ESM2AMP-MIC, a set of strain-specific MLP regression models that predict minimum inhibitory concentration (MIC) values against 75 bacterial and fungal strains, attaining an average R² of 0.572 and a mean Pearson correlation coefficient of 0.784. We applied ESM2AMP to metagenomic datasets from two transects in the Bohai Sea of China encompassing 81 samples across winter and summer seasons. The results revealed pronounced seasonal and spatial heterogeneity in AMP distribution: winter samples exhibited higher AMP abundance and stronger broad-spectrum antimicrobial potency, whereas summer-derived AMPs displayed greater target specificity. Strain-level MIC profiling further uncovered significant intra-species variation in susceptibility for clinically important pathogens such as Escherichia coli and Staphylococcus aureus. These findings demonstrate the power of protein language models in large-scale AMP mining from environmental metagenomes and provide candidate molecular resources for the development of novel antimicrobial agents.