AI and the World's Hard Problems · Life sciences · arXiv · 2026-10-20
HyperTaxa: Multiscale Genome Representations for Generalized Microbial Taxonomic Inference
Microbial taxonomic inference is difficult when many newly sequenced organisms lack reference coverage, and alignment or database tools cannot infer species absent from training. The authors propose HyperTaxa, which encodes whole genomes as multiscale k-mer transition fields and aligns them to a taxonomy-derived space, enabling zero-shot and open-set prediction. It outperformed pseudo-image and transformer-graph baselines on benchmark and disjoint-species evaluations, with the largest gains at higher taxonomic ranks.