Perseus: Lineage-Aware Refinement of Kraken2 Taxonomic Classification for Long Read Metagenomes

Area of Science:

• Metagenomics
• Bioinformatics
• Computational Biology

Background:

• Long-read sequencing advances microbial community analysis but faces challenges in accurate taxonomic classification.
• Existing k-mer classifiers like Kraken2 often over-assign labels, leading to high false positive rates, especially with novel microbes.

Purpose of the Study:

• To introduce Perseus, a novel framework for lineage-aware confidence estimation in taxonomic classification.
• To address the limitations of current methods in accurately classifying long metagenomic reads and assembled contigs.

Main Methods:

• Perseus models the spatial distribution and hierarchical consistency of k-mer evidence.
• It reframes classification as hierarchical confidence estimation using a convolutional neural network to refine taxonomic signals.
• The framework confirms, backs off, or abstains from assignments based on calibrated confidence scores.

Main Results:

• Perseus substantially reduces false assignment rates across simulated and real-world metagenomic datasets.
• It improves precision and lineage-consistent accuracy, particularly for long reads and assembled contigs.
• The method effectively distinguishes true taxonomic signals from spurious matches in novel microbiomes.

Conclusions:

• Perseus offers a robust solution for accurate taxonomic classification in long-read metagenomics.
• The framework enhances reliability by prioritizing correctness and lineage consistency.
• It integrates seamlessly with existing Kraken2 workflows.