Codon-aware multi-scale feature fusion for metagenomic sequence classification

Microorganisms exert profound influences on both natural ecosystems and human society. As the cornerstone of microbial research, metagenomics relies on the precise analysis of large-scale, multi-source genomic data. However, discriminating among diverse sequence types remains a formidable challenge. To address this, we present CamFi (Codon-aware Multi-scale Feature Fusion), a unified framework integrating overlapping nucleotide-triplet representations with multi-scale dilated convolutions for classifying prokaryotic chromosomes, eukaryotic chromosomes, plasmids, and viruses. Benchmark evaluations demonstrate strong mean F1 scores of 97.70% for eukaryotic chromosomes and 94.13% for plasmids. On the CAMI II marine dataset, CamFi achieved a weighted F1 of 82.65%, ranking second among six methods based on reported aggregate metrics. In a balanced four-class benchmark, CamFi attained a macro-F1 of 93.77%, compared with DeepMicroClass (90.03%) and the XGBoost stage of 4CAC (64.66%). These results establish CamFi as a competitive approach for unified metagenomic contig classification under the evaluated conditions.