Machine learning prediction of eukaryotic hosts for giant viruses

Abstract

Giant viruses (GVs; Nucleocytoviricota) infect diverse eukaryotes and are ecologically important across ecosystems. Although cultivation-independent sequencing has recovered tens of thousands of GV genomes from environmental samples, eukaryotic hosts are only known for a few isolates, leaving the host contexts of most GVs elusive. We developed GVHoP (Giant Virus-Host Predictor) that predicts eukaryotic hosts from GV genomes by integrating gene content and GV-eukaryotic sequence similarities. Trained on isolates with experimentally identified hosts, GVHoP achieved 97% accuracy in cross-validation and predicts hosts at three hierarchical levels of eukaryote classification. Functional analyses link top predictive features to various processes involved in virus-host interactions, including viral entry, replication, morphogenesis, and cellular metabolic reprogramming. Applying GVHoP to 7,897 GV metagenome-assembled genomes assigned hosts to 5,280 viruses and uncovered potential novel hosts across major GV orders that cannot be inferred from core-gene phylogenetic information. The predicted host composition clearly separates aquatic and terrestrial environments and distinct aquatic ecosystems. Together, GVHoP links viruses known only from nucleic acid sequences to their putative hosts across the eukaryotic tree of life, improves our understanding of GV-host interactions and their potential impacts on ecosystems, and paves the way for further ecological and functional studies of environmental GVs.