Machine learning models for delineating marine microbial taxa

Abstract

Abstract The relationship between gene content differences and microbial taxonomic divergence remains poorly understood, and algorithms for delineating novel microbial taxa above genus level based on multiple genome similarity metrics are lacking. Addressing these gaps is important for macroevolutionary theory, biodiversity assessments, and discovery of novel taxa in metagenomes. Here, I develop machine learning classifier models, based on multiple genome similarity metrics, to determine whether any two marine bacterial and archaeal (prokaryotic) metagenome-assembled genomes (MAGs) belong to the same taxon, from the genus up to the phylum levels. Metrics include average amino acid and nucleotide identities, and fractions of shared genes within various categories, applied to 14 390 previously published non-redundant MAGs. At all taxonomic levels, the balanced accuracy (average of the true-positive and true-negative rate) of classifiers exceeded 92%, suggesting that simple genome similarity metrics serve as good taxon differentiators. Predictor selection and sensitivity analyses revealed gene categories, e.g. those involved in metabolism of cofactors and vitamins, particularly correlated to taxon divergence. Predicted taxon delineations were further used to de novo enumerate marine prokaryotic taxa. Statistical analyses of those enumerations suggest that over half of extant marine prokaryotic phyla, classes, and orders have already been recovered by genome-resolved metagenomic surveys.

Key findings

• ▸Machine learning classifiers achieved balanced accuracy above 92% in identifying taxonomic relationships at genus to phylum levels.
• ▸Simple genome similarity metrics were found to be effective in differentiating between microbial taxa.
• ▸Gene categories involved in metabolism and other functions were identified as important predictors of taxonomic relationships.

Keywords

Identifiers

DOI

10.1093/nargab/lqaf090

Journal

NAR Genomics and Bioinformatics

Year

2025