Protein-language-model-based antibiotic resistance gene prediction and resistomes mining in cyanobacterial blooms

Antibiotic resistance genes (ARGs) are emerging microbial contaminants that accumulate and disseminate in freshwater ecosystems, yet their ecological dynamics during cyanobacterial harmful algal blooms (CyanoHABs) remain poorly characterized. Here, we developed ESMARG, a protein language model framework built on ESM1v, for integrated ARG identification, classification, and functional annotation. By incorporating abundant non-ARG sequences from real-world metagenomes, ESMARG achieved 0.990 precision, 0.837 recall, and 0.907 F1-score, outperforming alignment-based tools (BLAST, DIAMOND) and deep learning models (ARGNet, ARG-SHINE, PLM-ARG). Applying ESMARG to 21 cyanobacterial aggregate metagenomes from Lake Taihu spanning a full annual bloom cycle, we identified 145 ARGs across 15 drug classes and 6 resistance mechanisms, including 30 candidate ARGs. ARG abundance exhibited pronounced seasonal variation, peaking in summer (12.5 ± 8.68 copies per cell) when bloom intensity was highest. Procrustes analysis revealed strong coupling between resistome composition and microbial community structure, particularly cyanobacterial assemblages (matrix correlation = 0.62, p < 0.001). Structural equation modeling further revealed that cyanobacterial abundance strongly shaped bacterial community composition statistically (β = 0.87), while ARG abundance was jointly predicted by biochemical oxygen demand (β = −0.58) and bacterial replication-and-repair functions (β = 0.50). Water temperature and dissolved oxygen were identified as key environmental factors associated with these dynamics. These findings establish cyanobacterial blooms may serve as critical hotspots for ARG enrichment in Lake Taihu and provide a high-resolution methodological framework for monitoring resistome dynamics in bloom-impacted freshwater systems.