Nanopore Metagenomics Benchmark

Species Scorecard

Kraken2 species calls on a Nanopore run of a defined gut mock community (17 species, 21 strains), scored once on the held-out evaluation pool against the theoretical composition, with a sequencing-depth series and the confidence-threshold arms.

0.966Species F1, evaluation pool (14 TP · 1 FP · 0 FN)
14/17Species evaluable (≥ 20 expected reads)
1False positive: Homo sapiens (818 reads)
0.168Bray-Curtis distance to the theoretical composition

Observed against expected reads, per species

Kraken2 PlusPF-8, evaluation pool. A bar reaching the dashed line means the observed read count matched the theoretical composition. Grey bars are species with fewer than 20 expected reads, which could not be evaluated. Choose a sequencing depth to see how the picture changes (mean of 5 random subsamples; 1.0 Gb is the full pool).

Depth: Sort:
Evaluable species Not evaluable (< 20 expected reads)

Roseburia hominis (0.41) and Bifidobacterium adolescentis (0.22) fall below 1, and Limosilactobacillus fermentum (1.66) sits above it. Several causes are mixed here. L. fermentum reads are short (about 2 kb against 7 to 9 kb for most species), so its read share exceeds its DNA share. Exploratory work on the development pool found the Roseburia deficit already present in the mapped reads (the sequenced sample itself), whereas the classifier loses reads mainly for E. coli and B. adolescentis (the 8 GB database). The ratios are stable across depths. The expected composition is the theoretical one, because the sequenced kit lot could not be confirmed.

Confidence threshold: F1 on the evaluation pool

Kraken2 confidence thresholds were registered as arms, chosen on the development pool and scored once on evaluation. Nanopore reads rarely reach high k-mer confidence, so the default (0) was best.

At confidence 0.2 and 0.4 nothing is classified, so F1 is 0. Values from the registered confidence-arm summary.

View the repository on GitHub →