RNA-dependent RNA polymerases (RdRp) are crucial for replication of RNA viruses. They are also used as viral hallmark genes for evolutionary reconstruction and taxonomical classification.  Analyses of viral phylogenies  taxons can provide hypotheses about the origin and evolution of biological traits such as their host range in plant and fungus kingdoms. Recent discoveries, in particular cross-kingdom transmission events, highlight ecological and evolutionary links between plant and fungal viruses. Date estimation of these links became possible using the Prisoner of War model (PoW), a newly developed tree reconstruction method which accounts for timedependent rate effects over long timescales. However, the PoW model application requires a very accurate alignment, difficult to reach at high taxonomical ranks using traditional alignment algorithms (Clustal, Muscle, Mafft...). To overcome this issue, we leveraged both the structural protein conservation and the advancements in deep-learning algorithms. The reliable structural models predicted with Alphafold, an artificial intelligence program, allows for the detection of  similarities at deeper evolutionary timescales on the vast amount of sequence data derived from metagenomics studies and to infer improved structural alignments and more accurate phylogenies. Our interdisciplinary approach integrating structure modeling and date estimations on RdRps provided new insights on the very ancient evolutionary divergences between fungus and plant viruses and will contribute to optimize the classification of viral taxons and metagenomic sequences.