TaxDistill uses 500M-param GenomeOcean to boost metagenomic annotation F1 by 23%

Metagenomic taxonomic annotation—identifying which microbes DNA fragments come from—is crucial for understanding environmental and gut microbiomes. Traditional similarity-search methods (e.g., MMseqs2, BLAST) struggle with high microbial diversity and incomplete reference databases. Learning-based approaches like Taxometer attempt post hoc correction, but they inherit noise from the labels produced by those same search tools, degrading performance. To break this cycle, Rongye Ye and colleagues introduce TaxDistill, a knowledge distillation framework that leverages a large pre-trained genomic foundation model as a teacher.

The teacher, GenomeOcean (500M parameters), extracts deep semantic features from DNA sequences and generates soft labels based on confidence scores. These soft labels are distilled into a lightweight student network, effectively washing out the noise introduced by initial retrieval tools. Experiments on seven diverse CAMI2 datasets show consistent gains over baselines. On the Gastrointestinal dataset, TaxDistill boosts MMseqs2’s F1 score from 0.763 to 0.941, surpassing Taxometer. The paper demonstrates that distillation from a genomic foundation model is a reliable strategy for label correction in complex metagenomic analysis, offering a practical path to more accurate microbial profiling without requiring massive compute at inference time.