Metagenomic sequencing has dramatically expanded our view of the global virome, yet our ability to assess the biological relevance of novel viral sequences remains severely limited. Current approaches rely almost exclusively on sequence similarity to known viruses — a strategy that fails systematically for the vast, uncharacterized fraction of environmental viral diversity. Predicting which hosts a virus can infect is a critical step toward evaluating zoonotic risk, understanding transmission dynamics, and enabling actionable surveillance.
We present ViHostFinder, a universal viral host predictor built on a hierarchical multilabel classification framework using DNA language models. By operating directly on raw genomic sequences, ViHostFinder bypasses protein annotation — a step that remains unreliable for highly divergent viruses — making it broadly applicable across the virome. The model natively supports multilabel prediction, capturing the capacity of viruses to infect multiple host species and accounting for vector-mediated transmission. Hierarchical classification enables confident prediction at higher taxonomic levels when fine-grained labels are data-limited.
To handle the unequal distribution of sequence information across viral genomes — a key challenge for prediction on partial inputs — ViHostFinder incorporates mixture-of-experts and contrastive learning strategies, enabling robust inference even from fragmented metagenomic assemblies. Finally, the framework includes a metric to estimate host jump capacity, directly relevant to the emergence of novel viral threats. ViHostFinder will be freely available as a web server, designed with applicability in mind for the broader virology community.

Co-authors: Frank Møller Aarestrup, Henrik Nielsen, Philip Thomas Conradsen Clausen