What an independent benchmark found.

Source & scope: Prosser et al. (2026), "The Metabarcoding Analysis Pipeline (MAP)", bioRxiv. It is a preprint and has not been peer-reviewed, and it introduces the competing pipeline MAP, so its authors have an interest in their own pipeline performing well. spcfy was invited to comment before submission. Measurement is a single CO1 dataset, 10 arthropod bulk samples against single-specimen barcoding as ground truth. Only spcfy, MAP and mBRAVE return like-for-like BIN assignments, so those three are shown. Read it as one CO1 data point, not a verdict.

How the three pipelines compare on five metrics

Measured at the stringent filter level (≥0.001%), which is the setting spcfy recommends for production work. Of the five metrics, spcfy leads on one, Mantel R. The paper also reports a permissive and an unfiltered level; all three are in the table further down.

spcfy MAP mBRAVE

CCC (α-diversity)

higher is better

spcfy0.66

MAP0.66

mBRAVE0.77

Spearman ρ (α-diversity)

higher is better

spcfy0.89

MAP0.90

mBRAVE0.90

F1 (γ-diversity)

higher is better

spcfy0.62

MAP0.73

mBRAVE0.76

Mantel R (β-diversity)

higher is better

spcfy0.42

MAP0.33

mBRAVE0.35

Bray-Curtis distance (β-diversity)

lower is better

spcfy0.45

MAP0.35

mBRAVE0.34

Darker label = best of the three on that metric. spcfy's conservative filters keep β-diversity structure clean (Mantel R) but trim some low-abundance borderline detections along with noise, which is what pulls CCC, F1 and Bray-Curtis below MAP and mBRAVE. See the reasoning in full below the table.

Every metric at every filter level

Filter	Pipeline	CCC (α) ↑	Spearman ρ ↑	Mantel R ↑	BC-dist ↓	F1 (γ) ↑
Unfiltered	spcfy	0.29	0.71	0.23	0.57	0.59
	MAP	0.75	0.82	0.32	0.38	0.72
	mBRAVE	0.059	0.85	0.041	0.60	0.69
Permissive ≥0.0001%	spcfy	0.71	0.88	0.43	0.49	0.62
	MAP	0.81	0.87	0.34	0.36	0.76
	mBRAVE	0.37	0.92	0.15	0.44	0.77
Stringent ≥0.001%	spcfy	0.66	0.89	0.42	0.45	0.62
	MAP	0.66	0.90	0.33	0.35	0.73
	mBRAVE	0.77	0.90	0.35	0.34	0.76

Mint cells mark where spcfy is best of the three shown, Mantel R across all three filters. Full per-pipeline tables, including MetaWorks and QIIME2, are in the preprint's supplement.

Where spcfy stood out

β-diversity (Mantel R): consistently strong across all filters, the same conservative filtering that lowers recall keeps this clean.
Speed & usability: the paper explicitly calls spcfy fast and very user-friendly.
Once filtered, α-diversity (CCC) climbs sharply, from 0.29 unfiltered to 0.66–0.71.

Where spcfy scored lower and why

F1, CCC and Bray-Curtis read lower than MAP or mBRAVE, the recall-sensitive metrics.
spcfy's conservative filters cut some low-abundance borderline detections along with noise. That is spcfy's own reading of the result, not the paper's.
The tested version had no long-read support, the comparison is Illumina-only.

What this result means in practice

The trade is deliberate. The same filters apply to everyone's data, which is what lets results from different labs line up at the annotation layer. Two settings still belong to you: whether sample-wide singletons are kept, and the minimum and maximum length of your target region. What spcfy does not do is drop a fixed percentage of reads per sample. It used to, and stopped: the read counts in your OTU and ESV tables are the ones the pipeline actually worked with, and any subsampling is a decision you make downstream, on your own terms. Long-read support has since gone into production, so this benchmark is worth repeating on Nanopore data.

Compare: Overview · Feature Matrix · Cost · Benchmark · Alternatives