Résumé
Abstract RNA-dependent RNA polymerases (RdRps) are crucial for replication of RNA viruses and serve as key marker genes for defining deep taxonomic ranks and for understanding viral evolutionary history. Despite shared functions and conserved amino acid motifs, the high genetic diversity of RdRps complicates precise sequence comparisons across viral families, hindering accurate taxonomic classification of new species - especially important at the age of metagenomics. When available, three-dimensional (3D) RdRp structures can help address these challenges through structure-based alignments. However, such structures are scarce for understudied viruses infecting fungi and plants, preventing the investigation of their ecological and evolutionary links. In this study, we focused on the highy divergent order Sobelivirales. Using deeplearning structural modeling, we generated highly reliable 3D models on 20 representative viral species. Multiple structural alignment enabled the reconstruction of a robust phylogeny with improved quality and length. Based on this phylogeny, we proposed revisions of existing viral families and reclassified genera. Clade divergence dates were then estimated using the Prisoner of War model, which has previously revealed the ancient origin of the genus Sobemovirus. We provided here the first divergence time estimation between these plant and fungal viruses, dating back to 26.6 million years before present - significantly more recent than the divergence between their respective hosts. Our amino acid conservation analysis, validated on 99 other viral species, also identified molecular signatures of sobeliviral families and genera, which could help in future taxonomic assignment and diagnostic tools development. This interdisciplinary approach integrating structure modeling and date estimations offers new insights into the evolutionary divergence between fungus and plant viruses, with potential applications to other viral orders and families. Author summary RNA-dependent RNA polymerases (RdRps) are essential for RNA virus replication and serve as important markers for classifying viruses and understanding their evolution. However, with the rising popularity of metagenomics and discovery of viruses with high genetic diversity in RdRps, it is difficult to compare viral families and accurately classify new species. When available, 3D structures of RdRps can help overcome this challenge through structural alignments. In this study, we focused on the highly divergent order of the Sobelivirales, using deeplearning models to generate reliable 3D structures for 20 representative viral species. These structural alignments allowed us to build a more accurate viral phylogeny. Based on this finding, we proposed updates to existing viral families and genera within the Sobelivirales order. We also estimated divergence dates using a model that previously uncovered the ancient origins of Sobemovirus. Notably, we provided the first estimate of when these plant and fungal viruses diverged - around 26.6 million years ago, which is much more recent than the separation of their hosts. We also identified molecular signatures that are useful for future virus classification and diagnosis, with potential applications to other viral groups.