ESM2-Guided Context-Aware Annotation Completion Supplements Carbohydrate Metabolism Coverage in Silage Microbial Metagenomes

Abstract

Functional annotation gaps limit the interpretation of carbohydrate metabolism in silage microbiomes. We developed Context-Aware Annotation Completion (CAAC), a framework integrating ESM2 embeddings, genomic-neighborhood features, three-class classification, confidence-tiered neighbor voting, and Enzyme Commission (EC)-to-KEGG Orthology (KO) mapping. CAAC was applied to 21 metagenomes from uninoculated and Lacticaseibacillus paracasei-inoculated silages sampled before ensiling and at 7 and 90 days. Five-fold cross-validation yielded an F1-macro of 84.64% for negative, positive, and hard-sequence classification. Among 800,000 selected annotation-poor sequences, 545,671 Tier 1 or Tier 2 predictions passed the annotation-validity and EC-to-KO mapping criteria, of which 524,814 were eligible for sample-level annotation supplementation. After silage-focused filtering and KO–EC summarization, these predictions yielded 102 KO–EC features repeatedly detected across the silage metagenomes and increased coverage in 25 of 47 carbohydrate-metabolism pathways, mainly by recovering enzyme-level components related to starch and sucrose, cellulose and cellobiose, xylan and hemicellulose, and pectin and glucuronate metabolism. Taxon-linked analyses further revealed treatment- and stage-associated patterns in the taxonomic sources of the supplemented annotations. A database-derived temporal benchmark using the July 2025 CAZy release showed 94.94% Tier 1 family-level annotation-transfer consistency. CAAC extends the enzyme-level interpretation of under-annotated silage metagenomes, while the inferred assignments remain computational predictions requiring experimental validation.