Protein embeddings reveal a continuous molecular landscape of host adaptation in waterfowl parvoviruses

Viral adaptation across closely related hosts often proceeds through subtle molecular changes that escape detection by classical phylogenetic analyses. In waterfowl parvoviruses, we integrate AI-based protein language modeling, structural biophysics, and infection assays to reveal a continuous trajectory of host adaptation linking Goose parvovirus (GPV) and Muscovy duck parvovirus (MDPV). Protein embeddings of VP1 sequences reveal a smooth manifold bridging GPV and MDPV, which softens an apparent phylogenetic dichotomy into a graded molecular topology. Structural modeling identifies a flexible surface loop (residues 300-420) as a biophysical pivot. Along the embedding trajectory, this loop undergoes gradual conformational expansion and electrostatic neutralization, quantitatively linking embedding coordinates to capsid surface remodeling. Experimentally, a GPV-type isolate recovered from naturally diseased ducks replicated efficiently in duck embryos, duck embryo fibroblasts, and live ducklings, producing characteristic lesions. These results show that waterfowl parvoviruses evolve along a continuous molecular-electrostatic landscape in which cumulative structural adjustments enable cross-host infectivity. Our framework connects AI-derived molecular representations to biophysical mechanisms and biological function, supporting a model of viral host adaptation as a predominantly continuous process and providing a foundation for predicting cross-host potential in emerging viral systems.