Machine learning models for delineating marine microbial taxa.

Area of Science:

• Microbial genomics and bioinformatics
• Computational biology
• Marine microbial ecology

Background:

• The link between gene content and microbial taxonomic divergence is poorly understood.
• Existing algorithms for delineating novel microbial taxa above the genus level using multiple genome similarity metrics are lacking.
• Accurate microbial taxonomy is crucial for macroevolutionary theory, biodiversity assessments, and metagenomic discoveries.

Purpose of the Study:

• To develop machine learning classifier models for delineating microbial taxa from genus to phylum levels.
• To assess the utility of multiple genome similarity metrics in differentiating prokaryotic taxa.
• To enumerate marine prokaryotic taxa and estimate the recovery rate of higher taxonomic ranks.

Main Methods:

• Developed machine learning classifiers using average amino acid identity, average nucleotide identity, and fractions of shared genes.
• Applied models to 14,390 non-redundant marine bacterial and archaeal metagenome-assembled genomes (MAGs).
• Performed predictor selection and sensitivity analyses to identify key differentiating gene categories.

Main Results:

• Classifiers achieved balanced accuracy exceeding 92% at all taxonomic levels (genus to phylum).
• Genome similarity metrics effectively differentiate microbial taxa.
• Gene categories involved in metabolism (e.g., cofactors and vitamins) strongly correlated with taxon divergence.

Conclusions:

• Simple genome similarity metrics are robust differentiators for microbial taxa.
• Machine learning models provide a reliable framework for microbial taxon delineation.
• Over 50% of extant marine prokaryotic phyla, classes, and orders have likely been recovered in current genome-resolved metagenomic surveys.