Taxometer: Improving taxonomic classification of metagenomics contigs

Area of Science:

• Bioinformatics
• Computational Biology
• Metagenomics

Background:

• Current metagenomic classification relies on sequence similarity, which has limitations.
• Metagenomic binning utilizes both sequence and abundance data for clustering contigs.
• There is a need for improved methods to enhance taxonomic annotation accuracy.

Purpose of the Study:

• To introduce Taxometer, a novel neural network-based method for improving metagenomic contig taxonomic classification.
• To leverage contig abundance profiles and tetra-nucleotide frequencies for enhanced annotation.
• To provide a quality estimation for existing taxonomic classifiers.

Main Methods:

• Development of Taxometer, a neural network model.
• Integration of contig abundance profiles and tetra-nucleotide frequencies as input features.
• Application and benchmarking of Taxometer on short-read (CAMI2) and long-read metagenomic datasets.

Main Results:

• Taxometer increased the accuracy of species-level contig annotations by MMSeqs2 from 66.6% to 86.2% on CAMI2 datasets.
• Reduced incorrect species-level annotations by an average of two-fold for Metabuli, Centrifuge, and Kraken2 on the CAMI2 Rhizosphere dataset.
• Enabled benchmarking of taxonomic classifiers on complex long-read datasets lacking ground truth.

Conclusions:

• Taxometer significantly improves the accuracy and reliability of metagenomic taxonomic classification.
• The method enhances existing taxonomic annotation tools and provides quality estimates.
• Taxometer is a versatile, open-source tool applicable to various metagenomic data types and classifiers.