Genome architecture shapes the appropriate unit of comparison in small plant DNA viruses

Comparative viral genomics often assumes that a single unit of analysis can be applied across datasets. However, this assumption becomes difficult to justify when viral genomes may be monopartite or multicomponent, and when multicomponent genomes differ in whether their components are co-packaged or separately encapsidated. Here, we developed an architecture-aware, alignment-free framework to compare five plant DNA virus systems spanning monopartite and separately encapsidated multipartite architectures, including two-component and six-component systems: cacao swollen shoot virus complex (CSSV), banana streak viruses (BSV), Dioscorea bacilliform viruses, cassava mosaic begomoviruses (CMB) and banana bunchy top virus (BBTV). After curation of public GenBank records, 1,342 sequences were resolved into three analytical panels, including a 130-genome monopartite panel comprising CSSV and BSV and a 1,206-record componentized panel. In the monopartite panel, whole-genome compositional geometry robustly separated CSSV and BSV. A sliding-window mosaicity analysis further prioritized a subset of CSSV genomes with elevated internal discordance, consistent with mosaic-like compositional heterogeneity. The principal panel-level results were retained with canonical 5-mers, and mosaicity rankings were stable across 300-, 400- and 500-nt windows and concordant with an independent alignment-based similarity-profile benchmark. In the componentized systems, component identity formed a strong and reproducible axis of sequence organization in both CMB and BBTV. However, their isolate-level configurations differed: the complete CMB pairs examined here consistently selected different nearest external isolates, whereas complete BBTV sets showed more heterogeneous neighbour configurations and higher component-configuration tension. These results suggest that, within the systems examined here, the biologically informative unit of viral comparison may shift with genome architecture, from whole genomes in the monopartite panel to paired or distributed component configurations in CMB and BBTV. We therefore propose architecture-aware, alignment-free comparison as a cautious screening framework for identifying cases in which deeper phylogenetic or recombination-focused follow-up is most warranted.