VirRep: accurate identification of viral genomes from human gut metagenomic data via a hybrid language representation learning framework

Accurate identification of viral genomes from metagenomic data provides a broad avenue for studying viruses in the human gut. Here, we introduce VirRep, a novel virus identification method based on a hybrid language representation learning framework. VirRep employs a context-aware encoder and a composition-focused encoder to incorporate the learned knowledge and known biological insights to better describe the source of a DNA sequence. We benchmarked VirRep on multiple human gut virome datasets under different conditions and demonstrated significant superiority than state-of-the-art methods and even combinations of them. A comprehensive validation has also been conducted on real human gut metagenomes to show the great utility of VirRep in identifying high-quality viral genomes that are missed by other methods.