Does hierarchy help? flat vs. hierarchical classification of organisms from metabolic reaction (hyper-)networks

Metabolic networks encode a rich, measurable taxonomic signal, but whether this signal is sufficient to support accurate, large-scale hierarchical organism classification (and what it actually takes to exploit it) has not been systematically established. A large-scale supervised classification study of 5,621 organisms, each represented as a metabolic hypergraph and annotated with a complete four-level taxonomic path (from cellular organization to genus) derived from the KEGG BRITE taxonomy, demonstrates that reaction-level metabolic structure alone contains sufficient information to accurately recover the full taxonomic path for the large majority of organisms. Eighteen model--representation--classification configurations, spanning classical machine learning pipelines, kernel methods and Graph Neural Networks combined with flat and hierarchical classification strategies, are evaluated under a rigorous, multi-split, multi-metric protocol. Notably, this performance is attained by comparatively simple and computationally inexpensive methods, which match or exceed substantially more complex kernel- and Graph Neural Network-based alternatives, while flat classifiers perform on par with explicit hierarchical decompositions without requiring their additional architectural complexity. Prediction errors concentrate almost exclusively at the finest, most imbalanced taxonomic level, highlighting rare and fine-grained classes as the main remaining challenge.