MicroRAG: A Hybrid Language Model Retrieval-Augmented by Reference-Based Algorithms for Metagenomic contig Classification

MicroRAG, an reference-guided retrieval-augmented language model for microbial contig sequence classification. MicroRAG combines multi-scale convolutional feature extraction, a fragment-aware transformer to capture intra-contig dependencies, and a retrieval-augmented generation mechanism that incorporates external reference databases. Evaluated on a large-scale, temporally partitioned benchmark spanning viruses, plasmids, prokaryotes, and microeukaryotes, MicroRAG achieves 18–25% performance gains across accuracy, precision, recall, F1, and AUPR compared with state-of-the-art methods. This strength is particularly pronounced on short contigs and imbalanced communities, a capability attributed to its retrieval and representation components, both demonstrated as indispensable by ablation analyses. Applications to fecal microbiota transplantation and marine metagenomes further reveal MicroRAG’s ability to capture plasmid dynamics linked to dysbiosis and detect ecologically specific viral signals. By bridging deep representation learning with reference-based retrieval, MicroRAG establishes a scalable and interpretable paradigm for metagenomic sequence analysis.