Epsilon Archive for Student Projects

Abstract

Metagenomic Next Generation Sequencing is increasingly being adopted for clinical diagnostic
use as a valuable complement to traditional methods of pathogen detection. After major
improvements in accuracy, Oxford Nanopore sequencing represents a viable alternative to
Illumina. With longer read lengths, lower cost and shorter turnaround times, Nanopore can reduce
the time to diagnosis and improve patient outcomes. To realize this potential, sensitive taxonomic
classifiers with support for long-read data are needed. In this study we benchmark the performance
of different taxonomic classifiers on Nanopore-sequenced viral data from both mock and real
clinical samples. Custom databases for the Kraken2, DIAMOND, Metabuli, MetaCache and Sylph
classifiers were built from the same reference sequence data. Classifications were compared to
BLAST alignments as a gold standard and the classifiers were evaluated in terms of sensitivity and
precision.
Metabuli and MetaCache were the most sensitive across datasets and for different viruses, at
the cost of long processing times and high memory requirements respectively. Kraken2 showed
excellent precision but was the least sensitive. DIAMOND performed well on the mock data but
had lower species-level sensitivity on the shorter read length clinical data, likely reflecting the
limited specificity of protein-based classifiers. Sylph was highly computationally efficient and in
combination with Minimap2 performed well for most viruses, but was unable to detect some low-coverage genomes at default thresholds. Given enough memory, MetaCache may be the most
suitable classifier for a diagnostic workflow, possibly in combination with DIAMOND to leverage
the benefits of both DNA- and protein-based classifiers.