Quantitative assessment of NCLDV-host interactions predicted by co-occurrence analyses

Nucleocytoplasmic DNA viruses (NCLDVs) are highly diverse and abundant in marine environments. However, knowledge of their hosts is limited because only a few NCLDVs have been isolated. By taking advantage of the rapidly increasing metagenomic data, in silico host prediction approaches are expected to fill this gap between known virus-host relationships and the true but largely unknown amount of NCLDVs. In this study, we built co-occurrence networks between NCLDVs and eukaryotes using the Tara Oceans metagenome and metabarcoding datasets to predict virus-host interactions. Using the positive likelihood ratio to assess the performance of host prediction for NCLDVs, we demonstrated that co-occurrence approaches can increase the odds of predicting true positive relationships four-fold compared with random host predictions in the high-weight region (weight > 0.4). To refine the host predictions from high-dimensional co-occurrence networks, we employed a recently proposed phylogeny-based method, Taxon Interaction Mapper, and showed that Taxon Interaction Mapper further improved the prediction performance eight-fold using weight cut-off filtration (> 0.4). Finally, we inferred virophage and NCLDV networks that further corroborated that co-occurrence approaches are effective for predicting NCLDV hosts in marine environments. ImportanceNCLDVs infect a wide range of eukaryotes. Their life circle is less dependent on hosts than other viruses. However, our understanding of NCLDV-host systems is highly limited because few of these viruses have been isolated. Co-occurrence inference is a possible way to predict virus-host interactions. For the first time, we quantitatively assessed the effectiveness of co-occurrence analyses for NCLDV host prediction. We also improved the performance of these co-occurrence analyses with a phylogeny-guided filtering method and produced a list of candidate hosts for three NCLDV families.