Phage evolutionary relationships emerge from protein language model-based proteome representation

Viral taxonomy is a challenging task due to the propensity of viruses for recombination and the lack of universal gene markers. As a result, recent ICTV updates increasingly rely on multiple tools for taxonomic ranking, with a growing emphasis on proteome-based clustering approaches. At the same time, the rapid expansion of viral datasets presents new challenges in organizing, analysing, and discovering phage relationships at scale. To address these challenges, we introduce hierarchical viruses, a framework for comparative genomics of bacteriophages that leverages protein Language Model (pLM) embeddings to generate proteome-wide vector representations of phages. Clustering the vector representations of 24 362 phages from the curated INPHARED dataset reveals a multi-scale hierarchical organization of phages. This hierarchy aligns with current ICTV taxonomic rankings at the genus and subfamily levels, with an adjusted mutual information score greater than 0.9 for both, in the Herelleviridae family, demonstrating that pLM-based proteome representations can effectively capture evolutionary relationships without relying on multiple sequence alignments. The framework builds the basics towards vectorial phage datasets that encode evolutionary information, thus allowing discovery of phage relationships at scale.