Species-level k-mer evidence¶

Read this after Choosing confidence and threshold

Bracken's depth-scaled read threshold, described on that page, has the largest effect on precision in shotgun mode; set it first. This gate catches one narrow kind of error that the threshold cannot detect. It is off by default; see Why the default is enabled: false.

The gate is a separate setting from Bracken's read threshold (-t) and does not change the k-mer length. It counts the distinct minimizers of a species found in the sample. See the four stringency settings at a glance.

Kraken2's read count tells you how many reads it chose to label with a taxon. It does not tell you how much sequence unique to that organism actually turned up in the sample — and those are different questions.

A species that is genuinely present contributes many different species-specific sequences.
A spurious call is usually a small set of sequences — conserved, repetitive, or shared with a relative that really is abundant — matched over and over.

In a read count the two look identical. Running Kraken2 with --report-minimizer-data adds the number that separates them. MetaFlux always does.

The two extra columns¶

The report becomes eight columns instead of six, with two inserted before the rank code:

Column	Contents
4	Number of minimizers in the read data associated with this taxon
5	An estimate of the number of distinct minimizers in the read data associated with this taxon

Column 5 is an estimate, not an exact count — Kraken2 uses a HyperLogLog sketch rather than storing every minimizer. Both columns are clade-cumulative, like the read counts.

What "associated with this taxon" actually means

The columns do not count the minimizers carried by the reads assigned to the taxon.

Kraken2 attributes each minimizer occurrence to whichever taxon that minimizer maps to in the database (its lowest common ancestor), pooled over every read in the sample, regardless of where each read finally got assigned.

This is visible in any report: taxa with two or three assigned reads can carry hundreds of thousands of minimizer counts. In SRR39106521, one of the test samples described below, the genus row Streptomyces has 3 clade reads and 343,359 total minimizers, about 500 times the 690 k-mer positions that three pairs of 149-base reads can hold. Counts that high cannot come from the assigned reads alone; they are expected when minimizers are attributed by LCA.

So column 5 for a species means: how many distinct database minimizers specific to that species were seen anywhere in this sample. That is a statement about how much species-specific sequence is present — the minimizer counterpart of what KrakenUniq calls unique k-mers, so the published thresholds below are a fair starting point; how well they convert to minimizers is discussed below. It is not a genome-coverage measure, and describing it as "how much of the genome the reads covered" would be wrong.

The flag itself changes no classification

--report-minimizer-data is purely additive reporting: with or without it, Kraken2 writes a byte-for-byte identical per-read output and the same values in every shared report column.

Reading the numbers¶

Two species from SRR39106521, with similar read counts:

Species	Reads	Distinct minimizers	Total minimizers	Duplicity
Pseudomonas aeruginosa	4,775	68,276	221,570	3.2
Pigmentiphaga sp. CHJ604	4,129	929	226,726	244.1

The two also have almost the same total minimizer count. The first is supported by 68,276 distinct species-specific minimizers. The second is supported by 929, each one seen about 240 times.

Duplicity (total ÷ distinct) is the most directly readable of the three numbers, because it does not require knowing the sequencing depth. It also barely tracks abundance: across the 2,418 species rows in the Kraken2 reports of the four test samples, Pearson r against log reads is +0.17 for log duplicity and +0.71 for the log distinct count. That makes it the better number to eyeball. It is nonetheless a worse filter (see What does not work), so MetaFlux reports it and thresholds the distinct count.

What MetaFlux writes¶

Every shotgun run produces these, per sample, whether or not the gate is switched on:

02.classification/kraken2/{sample}_species_evidence.tsv   one row per taxon at the Bracken rank
02.classification/kraken2/{sample}_report_gated.txt       the report with failing taxa removed
stats/kmer_evidence/{sample}_evidence.png                 reads vs distinct minimizers, log-log

The gated report is written even when the gate is off, precisely so you can diff it against the Kraken2 report and see what enabling the gate would do to your data before trusting it.

The evidence table carries, per taxon: taxid, name, rank, clade and direct reads, total and distinct minimizers, duplicity, minimizers per read, percent of classified reads, and the verdict with its reason.

The gate¶

shotgun:
  kmer_evidence:
    enabled: false                     # report only; pipeline unchanged
    min_distinct_minimizers: 300       # the floor: a whole number, or auto (see below)
    min_reads: 0                       # extra read floor, applied together with the above
    never_prune_above_fraction: 0.01   # safety rail

Where it acts¶

kraken2  →  kmer_evidence  →  bracken  →  finalize_otu_table

The gate sits between Kraken2 and Bracken. It does not replace or skip Bracken — it hands Bracken a cleaned report instead of the raw one. Because the correction happens upstream, it flows all the way through: otu_table.biom and otu_table.tsv become the corrected tables. There is no separate filtered file to reconcile.

Switching the gate off restores the previous behaviour exactly: Bracken reads Kraken2's own report and nothing downstream changes.

How a taxon is removed¶

Two details must be handled correctly; otherwise Bracken redistributes the reads of the pruned report wrongly, without an error.

Whole subtree, never a bare row. Bracken rebuilds the taxonomy tree from the indentation of the report, attaching each row to the last row one level above it. Delete a species row and leave its strain (S1) children behind, and those children get re-parented onto whichever species happens to precede them, whose read count is then inflated without any warning. So when a taxon fails, everything indented beneath it goes too.

The reads are credited to the parent, not discarded. A pruned taxon's whole clade total is added to its parent's own-node count, so Bracken redistributes those reads among the siblings that passed. The parent's clade total is untouched, so the report's arithmetic still balances: for every node, clade == direct + sum(children clade), and no column other than a parent's direct count ever changes.

Two things this does not guarantee

Not all reads land on close relatives. The rationale for crediting the parent is that a spurious call usually borrows evidence from a neighbour. That holds for most of the reads that move, not all: part of them cross genus or family boundaries.

Some reads are still lost. If every species under a genus fails, Bracken has nothing left at the target rank to redistribute to, and drops those reads; the same happens one level up when every species of a family or an order fails. At the default floor of 300 that is a small fraction of the sample: on the test dataset the table lost at most 0.09 % of its read pairs. With auto it was 2.7 to 15.1 % of the same tables, because a whole order failed in two of them (see Limits of the rule). Crediting the parent is much better than deleting rows outright, but it is not conservation.

Choosing a threshold¶

Three published sources give thresholds for this kind of filter, all expressed in KrakenUniq's unique k-mers:

Source	What it reports
KrakenUniq (Breitwieser, Baker & Salzberg, 2018)	On a synthetic mix of reads from bacterial isolates (1–34 million read pairs), the best threshold (highest recall at a false discovery rate of 5 % or less) rose by about 2,000 unique k-mers per million reads sequenced: about 3,700 at 1 million, 21,000 at 10 million. For ten patient biopsies, mostly host DNA, a fixed 1,000 unique k-mers together with at least 10 reads removed the background calls and kept every confirmed pathogen.
Oskolkov (2026), Front. Microbiol. 17:1603339	Simulated ancient metagenomes (fragments of about 42 bases, 0.1–0.7 million reads): about 1,000 unique k-mers gave the best F1, and a read-count filter added nothing. The best value rose with depth, by about 200 unique k-mers per 100,000 reads, a trend the paper calls relatively weak. Run on the same data, Kraken2 needed two to three times more distinct minimizers than KrakenUniq needed unique k-mers.
aMeta (Pochon et al., 2023)	Defaults of 1,000 unique k-mers and 200 assigned reads, set from more than 1,200 ancient libraries; users are invited to try 500–1,500 k-mers depending on depth. The 2026 study above comes from the same group and reuses aMeta's simulated data.

Published thresholds count k-mers; Kraken2 reports minimizers

In plain terms. Kraken2 reads a sequence as overlapping 35-base k-mers, but it does not look each one up. For every k-mer it picks one 31-base piece inside it, the minimizer, and stores and counts only that. Neighbouring k-mers usually pick the same piece, so a region of a genome gives about three times fewer distinct minimizers than distinct k-mers. The studies above counted unique k-mers and settled on about 1,000. In a Kraken2 report the same amount of sequence shows up as about 333 distinct minimizers. The default floor, 300, is that figure rounded down, because the conversion is only approximate. It was not tuned on MetaFlux's own test data, and it is a low, cautious value rather than an exact equivalent. When Oskolkov (2026) ran both tools on the same simulated data, the best Kraken2 floor was two to three times higher than the best KrakenUniq one, where the conversion predicts three times lower, and MetaFlux's own benchmark finds useful floors in the thousands (see Scaling the floor with depth).

The arithmetic, for a database built with other settings. How many k-mers share a minimizer depends on the database's k and l. Each k-mer of length k spans w = k − l + 1 candidate minimizer positions of length l, and the standard expected density of selected minimizers is 2 / (w + 1). For the usual nucleotide settings (k = 35, l = 31) that is w = 5 and a density of 1/3 — roughly one distinct minimizer per three k-mers.

So, counting sequence alone, a published 1,000 unique k-mers corresponds to about 333 distinct minimizers here.

Check k and l on the first line of your database's inspect.txt and recompute if they differ. (The 2/(w+1) density is the expectation for a random minimizer scheme; Kraken2's spaced-seed masking shifts it slightly, so treat 333 as an approximate conversion, not an exact one.)

What it does on real data¶

The numbers below come from the four metagenomes of BioProject PRJNA1477390 (runs SRR39106521–SRR39106524), microbial consortia enriched on naproxen from wastewater treatment plants (see Shotgun test dataset). They were processed with the MetaFlux v2.5.0 defaults (Kraken2 confidence 0.15, Bracken threshold: auto) and PlusPF k2_pluspf_20260626, once with the gate off and once with it on at the default floor of 300. The floor is min_distinct_minimizers: the number of distinct minimizers a species needs in the Kraken2 report to stay in the table.

	SRR39106521	SRR39106522	SRR39106523	SRR39106524
Species in `otu_table.tsv`, gate off	99	80	116	120
Species, gate on (floor 300)	84	66	94	104
Genera, gate off → on	39 → 34	31 → 28	46 → 39	52 → 44
Read pairs held by the removed species	4.9 %	1.3 %	5.8 %	1.8 %
Change in total read pairs	−0.01 %	+0.01 %	−0.01 %	−0.09 %
Genomes detected by sylph in the same reads	10	9	14	16
Species above 1 % of `otu_table.tsv`, gate off	21	10	23	12
Abundant members counted from the 16S rRNA gene, no reference	7	6	4	10

The gate removes 14 to 22 species per sample and almost no reads. The reads of a removed species go back to its parent taxon, and Bracken shares them among the relatives that passed.
The removed species are not only rare ones. Bracken spreads the reads that Kraken2 could only place at genus level or above over the species below, and it had given the removed species 10 to 27 times more reads than Kraken2 assigned to them. In every sample one of the 20 most abundant species is removed. The largest case is Pseudomonas sp. JS425 in SRR39106521: 372 read pairs from Kraken2, carrying 182 distinct minimizers, raised by Bracken to 69,992 read pairs, the eighth most abundant species in the sample. Its read count passes the Bracken threshold; only the distinct-minimizer count shows how little species-specific sequence is behind it (more below).
Fewer species is the right direction for these samples. sylph, run by MetaFlux in both mode against gtdb-r232-c200-dbv1 and fungi-refseq-2025-10-11-c200-dbv1, detects 9 to 16 genomes per sample; it misses organisms without a closely related reference genome. A count that needs no reference database comes from the 16S rRNA gene fragments in the reads, extracted with Metaxa2 2.2.3: fragments from the same part of the gene are compared with one another and grouped at 97 % identity. It sees only the more abundant members, roughly those above 1–2 % of the 16S gene copies, and finds 4 to 10 per sample, mostly from the same families that sylph and Kraken2 report. For that abundant part of the community Kraken2 reports 10 to 23 species, those above 1 % of its table, because it splits the reads of one organism over several related species.
The gate does not bring the list down to that size. Most of what remains is a database problem. The most abundant organism in SRR39106521 and SRR39106523, Pusillimonas noertemannii, has no genome in PlusPF. Its reads are reported under several Achromobacter species, which together hold more than half of each of these two tables; the largest of them have tens of thousands of distinct minimizers (see the warning on missing organisms in Databases). A k-mer floor cannot correct that.

The same four samples at other floors: species left in otu_table.tsv, and the share of the gate-off table held by the species that the floor removes.

Floor	SRR39106521	SRR39106522	SRR39106523	SRR39106524	Read pairs held by removed species
none (gate off)	99	80	116	120	0
200	92	73	105	112	0.9–4.7 %
300 (default)	84	66	94	104	1.3–5.8 %
1,000	55	44	61	80	8–21 %
3,000	29	26	31	54	13–42 %

A lower floor keeps the species with 200 to 300 distinct minimizers, 7 to 11 more per sample. These samples have no known composition, so they cannot show whether those species are real. A higher floor removes many more species, and those species hold a large share of the table: at 3,000, 13 to 42 % of the read pairs. Part of them moves to other species and part leaves the table; how the two parts split was measured with auto, whose floors on these samples run from 1,673 to 3,458 (see Limits of the rule, below). On benchmark communities of known composition, a fixed floor of 10,000 already lowered some samples' F1 by removing true species, and lower fixed floors did the same once shallower sequencing was modelled (3,000 at 30 % of the depth, 1,000 at 10 %). A floor scaled to the depth of the sample avoids removing true species that way; see the next section.

Scaling the floor with depth¶

A species that is really present leaves more distinct minimizers the deeper a sample is sequenced, and so, up to a limit, does a spurious call that borrows sequence from an abundant relative. A fixed floor is therefore lenient on deep samples and strict on shallow ones. KrakenUniq and Oskolkov (2026) reported the same for unique k-mers (table above), but neither gives a depth rule for Kraken2.

MetaFlux's benchmark was used to fit one. It covers 42 samples of known composition: one real mock community (ZymoBIOMICS D6331, SRA run SRR29240428, BioProject PRJNA1118346) and 41 simulated CAMI II samples from the marine, rhizosphere and strain-madness sets (see Benchmark data), with 2.6–6.5 million classified read pairs each, classified with three Kraken2 indices: k2_pluspf_20260626 (at confidence 0.15 and 0.20), k2_gtdb_genome_reps_20250609 (GTDB r226 representative genomes) and k2_NCBI_reference_20251007 (one genome per NCBI species). A floor of one distinct minimizer per 1,000 classified read pairs gave the best result that was safe everywhere:

species-level F1 rose by 0.067 on average (0.036 with PlusPF), against 0.0025 for the fixed default of 300 (0.0003 with PlusPF);
it removed 2,291 of 13,117 false species calls and 6 of 13,419 true ones (none with PlusPF), and lowered no sample's F1;
it stayed safe when shallower sequencing was modelled, down to 1 % of the reads, while every fixed floor tested, from 300 to 10,000, lowered some samples' F1 at some depth.

This is the same order of magnitude as KrakenUniq's slope (about 670 distinct minimizers per million read pairs once converted), but the number itself comes from MetaFlux's benchmark.

`min_distinct_minimizers: auto`¶

auto applies this rule for you, sample by sample:

shotgun:
  kmer_evidence:
    enabled: true
    min_distinct_minimizers: auto

After Kraken2 has run, rule kmer_evidence works out each sample's floor from that sample's own Kraken2 report:

classified pairs   clade reads of the root row
human pairs        clade reads of the Homo sapiens row (0 if the report has none)
floor              (classified pairs − human pairs) / 1,000, at most 6,500

Classified read pairs, human excluded	Floor set by `auto`
150,000	150
500,000	500
1 million	1,000
2 million	2,000
5 million	5,000
6.5 million or more	6,500

One floor per sample. With a number, MetaFlux applies that one floor to every sample of the run. With auto, each sample is judged against its own floor, so a deep sample gets a higher floor than a shallow one in the same run. The four samples of the test dataset get floors of 1,673 to 3,458.
A species stays when its count reaches the floor: distinct minimizers ≥ the unrounded floor (1,672.619 in the example below), as in the benchmark, and it must also clear min_reads. The 1 % safety rail works as with a number; its share is of all classified reads, human ones included.
Human reads are left out. PlusPF contains the human genome and host removal is off by default, so human reads would raise the floor without adding any microbial evidence. This is MetaFlux's addition to the rule: in the benchmark, the host reads left in a sample counted as classified. The row is recognised by its taxid (9606) and its name, so a database whose taxids are not NCBI's (a GTDB-based one) is not affected. Reads of other hosts are not subtracted. With PlusPF they mostly stay unclassified; a database that holds other host genomes classifies them, and they then raise the floor: remove the host first (Decontamination) or set a number. The Homo sapiens row itself is still judged against the floor: in all four samples of the test dataset it fails (605 to 1,743 distinct minimizers) and leaves the table with its read pairs, where at 300 it stays.
The cap of 6,500. The benchmark stops at 6.5 million classified pairs. A floor that keeps growing with depth eventually exceeds what a real species with a small species-specific part of its genome can reach; in the model this began at about 10 million pairs.
Shallow samples. Below about 500,000 classified pairs the scaled floor gains little, and below 300,000 auto sets it lower than the default of 300. That is intended: in benchmark samples thinned to 26,000–195,000 classified pairs, the fixed floor of 300 lowered some samples' F1 and the scaled floor lowered none. On the one benchmark sample really run at 0.76 million classified pairs, auto (floor 759) and 300 removed the same three false species (below).
Where to see the floor. Each sample's log, logs/kmer_evidence/{sample}.log, shows how its floor was worked out:
```
[kmer_evidence] auto floor: 1673810 classified read pairs, less 1191 Homo sapiens read pairs (taxid 9606): 1672619 / 1000 = 1672.619. A taxon at rank S needs at least 1673 distinct minimizers.
```
The evidence table gives the floor in its reason column (for Sphingobium wenxiniae in that sample, distinct minimizers 1664 < 1673 (auto)), and the plot draws it, labelled (auto). Both show it rounded up: the counts are whole numbers, so "at least 1,672.619" and "at least 1,673" are the same test.

To size a number by hand instead, for instance to judge every sample of a run against the same floor, divide the classified read pairs of the shallowest sample, less its human read pairs, by 1,000. Each sample's classified count is in stats/read_tracking.txt (column classified), and its human read pairs are in the Homo sapiens row of its Kraken2 report. The deeper samples are then judged against a lower floor than auto would give them. Where their extra depth is spread over the whole community, that lower floor is more lenient than their depth warrants; where the extra depth sits on one dominant organism, whose reads raise auto's floor for the other species without adding to their evidence (see the end of this page), the lower floor is the fairer one. On the test dataset the single floor of 1,673 keeps 41, 33, 45 and 68 species where auto keeps 35, 25, 45 and 61.

Why `auto` is not the default¶

auto is opt-in, and the default stays 300:

Its gain rests mostly on one real mock community. Three quarters of the F1 gain came from the ZymoBIOMICS D6331 sample; on the three simulated CAMI II sets (41 samples) the mean gain per set was 0.01–0.03.
Part of its safety was modelled, not measured. For the benchmark samples the shallower depths were modelled from the observed counts, not sequenced. One of them was then subsampled to three shallower depths and run for real with PlusPF, and behaved largely as the model predicted (below); the other 41 rest on the model. No benchmark sample is deeper than 6.5 million classified pairs, which is why there is a cap.
It changes the table much more than 300 does, and read pairs leave the table: 2.7 to 15.1 % of the read pairs on the test dataset, against at most 0.09 % at 300, most of them from an organism with no close reference genome (the last point under Limits of the rule).

Measured at known depths and on deeper samples¶

What auto does on samples sequenced at known depths, and on samples deeper than any in the benchmark:

Measured at four depths. CAMI II marine sample 0, one of the benchmark samples, was subsampled to 2, 5 and 10 million read pairs and run together with its full 16.6 million, with the settings of its benchmark run (PlusPF k2_pluspf_20260626, Kraken2 confidence 0.15): once with the gate off, once at 300 and once with auto. After quality filtering, Kraken2 classified 0.76, 1.90, 3.79 and 6.31 million pairs. Species were scored as in the benchmark, against the sample's 252 gold-standard species:

Classified read pairs	0.76 M	1.90 M	3.79 M	6.31 M
Floor set by `auto`	759	1,896	3,788	6,309
True / false species, gate off	217 / 71	216 / 69	216 / 68	216 / 68
True / false species, 300	217 / 68	216 / 67	216 / 66	216 / 67
True / false species, `auto`	217 / 68	216 / 64	216 / 63	216 / 61
Recall, gate off, 300 and `auto`	0.861	0.857	0.857	0.857
Precision, gate off → 300 → `auto`	0.754 → 0.761 → 0.761	0.758 → 0.763 → 0.771	0.761 → 0.766 → 0.774	0.761 → 0.763 → 0.780
Species-level F1, gate off	0.804	0.804	0.806	0.806
F1 change, 300	+0.0045	+0.0030	+0.0030	+0.0015
F1 change, `auto`	+0.0045	+0.0076	+0.0076	+0.0107
True / false species, `auto`, as the depth model predicted	216 / 66	216 / 64	216 / 63	the model's starting point

No true species was removed, at any depth, by either floor: recall is the same with the gate off, at 300 and with auto, at every depth. Relative to its floor, the true species closest to removal was Aliivibrio wodanis at full depth, 7,263 distinct minimizers against 6,309 (1.15 times the floor); at 0.76 million the weakest true species, Synechococcus sp. WH 8020, had 1,186 against 759.
The depth model held where it was really tested. The model starts from the full-depth run and thins its calls to a shallower depth, so the 6.31 M column is its starting point, not a test, and the two middle depths are the real tests: at 1.90 and 3.79 million it gave the numbers of true and false species of auto exactly, and the true species kept the distinct minimizers it predicted for them (median within 1 %). At 0.76 million it was off in two ways. Species near Bracken's read threshold come and go by chance when reads are subsampled (six crossed it at 0.76 million, one at 1.90), which the model cannot see and no floor can correct; and three false species had 1.6 to 3.6 times fewer distinct minimizers than the model gave them, so at that depth 300 removed as many false species as auto did.
The read pairs in the table hardly changed: by less than 0.05 %, with either floor, at every depth.

This is one simulated sample on one index: it measures the depths the benchmark only modelled, for this sample, and says nothing about the other 41 benchmark samples.

A deep mock community. MBARC-26 is a mock of 23 bacteria and 3 archaea with known genomes (Singer et al., 2016), here in the library sequenced by Hardwick et al. (2018) (SRA run SRR6392275). With PlusPF k2_pluspf_20260626 it has 51.9 million classified read pairs, eight times the deepest benchmark sample, so auto sets the cap, 6,500; without the cap the floor would be 51,851. The gate was run on the existing Kraken2 report and Bracken was not run again, so what follows is which species each floor keeps, not how the read counts move.

Floor	Members removed, of 26	Other species rows of the report kept, of 729
300	Nocardiopsis dassonvillei	105
`auto` (the cap, 6,500)	N. dassonvillei	3
Without the cap (51,851)	N. dassonvillei, Escherichia coli	1

N. dassonvillei has 15 read pairs from Kraken2 in this library and 264 distinct minimizers, so every floor removes it from the report: the one true member a floor removed in these tests, and a very thin one. E. coli K-12 has 713 read pairs and 9,450 distinct minimizers, a member that is there and that a floor growing without limit would remove from the report; the cap keeps it.
The final table keeps the same species at every floor, the same as with the gate off: Bracken's read threshold at this depth (-t 2593) already leaves out both thin members, and the other 24 members, and Lysobacter enzymogenes, which is not in the mock (154,092 distinct minimizers), pass every floor.
The library has 2×125 bp reads, shorter than the 2×150 bp the rule assumes (see Limits of the rule), so the floor is relatively stricter here than on the benchmark; it still removed no member that Bracken keeps.

Deep, diverse soils. Seven forest soils from the Schönbuch, Germany (BioProject PRJNA717813; the O, A and B horizons of a Cambisol and the O and A horizons of a Podsol and a Stagnosol), with 50 to 108 million read pairs after quality filtering, of which PlusPF k2_pluspf_20260626 classified only 4.0 to 8.6 million. auto set floors of 3,953 to 6,500, the cap in four of them. Their composition is not known, so this shows how much auto removes, not whether it is right. As for the mock, the gate was run on the existing reports and Bracken was not run again.

Most species rows of the report go, few species of the table. auto removed 59 to 76 % of the species rows of each Kraken2 report (300: 9 to 13 %); those rows held 0.13 to 0.22 % of the classified reads. Of the 39 to 69 species per soil in the final table with the gate off, it removed none in three soils and one to three in the other four, all of them kept at 300: Kribbella sp. NBC_01484 in four soils (425 to 856 read pairs from Kraken2, 2,300 to 3,703 distinct minimizers) and two Bradyrhizobium species in one.
Those species held up to 5.6 % of a table, and that is the most a table can lose. Bracken had raised Kribbella sp. NBC_01484 from its 425 to 856 read pairs to 67,000–172,000; the three species removed from one soil held 5.6 % of its table, the one species removed from each of the other three 1.1 to 2.6 %. A species that passes the gate never loses read pairs when its relatives are removed, so the loss is at most this share; how much of it moves to relatives and how much leaves the table was not measured.
The cap changed no species of the table. Without it the four capped soils would get floors of 7,007 to 8,607, and 42 to 232 more species rows of their reports would go, none of them in the final table.

Limits of the rule¶

These hold with auto and with a number sized by hand:

It rests on limited data (see Why auto is not the default).
Do not go above 1.5 per 1,000 pairs when you size a number by hand; auto uses 1 per 1,000. At 1.5 rare true species start to be removed (92 of 13,419 true calls), and from 2 per 1,000 some samples' F1 falls even at full depth. On the CAMI II sample measured at full depth (above), 1.5 per 1,000 would set the floor at 9,463 and remove a true species, Aliivibrio wodanis (7,263 distinct minimizers).
Do not go above 6,500, by hand either, for the reason auto stops there (the cap, above).
Host reads. auto leaves out the Homo sapiens read pairs only. When you size a number by hand, subtract the read pairs Kraken2 assigned to the host (for human, the Homo sapiens row of the report), or remove the host first (Decontamination).
Library layout. The rule assumes paired 2×150 bp Illumina reads with little mate overlap and few duplicates. Shorter reads, single-end data, overlapping mates and duplicate-rich libraries give fewer distinct minimizers per read pair, so the same floor is stricter for them; use a lower number rather than auto.
The floor follows the depth of the whole sample, not the reads of each species. In a sample dominated by one organism, the other species face a higher floor without having more reads of their own; an example is at the end of this page.

It changes the table more than 300 does, and read pairs leave the table. On the test dataset above (1.7–3.5 million classified pairs), auto keeps 25–61 species per sample, against 66–104 at 300, and 2.7–15.1 % of the read pairs leave the table, against at most 0.09 % at 300:

	SRR39106521	SRR39106522	SRR39106523	SRR39106524
Classified read pairs	2,268,774	3,458,847	1,673,810	2,177,429
Homo sapiens read pairs, left out	1,843	1,721	1,191	1,451
Floor set by `auto`	2,267	3,458	1,673	2,176
Species in `otu_table.tsv`, gate off	99	80	116	120
Species, 300	84	66	94	104
Species, one floor of 1,673 for every sample	41	33	45	68
Species, `auto`	35	25	45	61
Genera, gate off → `auto`	39 → 12	31 → 10	46 → 18	52 → 23
Read pairs held by the species `auto` removes	29 %	13 %	33 %	14 %
Change in total read pairs, 300	−0.01 %	+0.01 %	−0.01 %	−0.09 %
Change in total read pairs, `auto`	−15.1 %	−11.1 %	−4.2 %	−2.7 %

Of the read pairs the removed species held, 15 to 87 % went to species that stayed; the rest left the table, because no species of their genus, family or order was left to take them.
Most of what left belongs to one order, Sphingomonadales. In SRR39106521 and SRR39106522 every one of its 24 and 23 species fails (the strongest, Novosphingobium pentaromativorans, has 2,174 and 2,156 distinct minimizers against floors of 2,267 and 3,458), and the whole order, 16 and 11 % of the two tables, is dropped; the rest of each table changes by about 1 % or less (+19,800 and −854 read pairs). Where one species of the order passes, it collects most of the order's reads: in SRR39106523 N. pentaromativorans goes from 40,710 to 283,996 read pairs, on about the same distinct-minimizer count as in the two samples that drop it.
The order seems to be a real member of these communities without a close reference genome, not a spurious call. That reading rests on two indirect signs, not on a measurement: the 16S rRNA gene fragments in the same reads, classified by Metaxa2 against SILVA, put 13–21 % of them in the order, and sylph finds no genome of the order in any sample. Kraken2 then spreads its reads over many named species, none with much species-specific sequence, and the floor removes them all.
auto also removes Homo sapiens from all four samples (605 to 1,743 distinct minimizers, below its sample's floor in all four); at 300 the row stays. A user who reads human contamination off otu_table.tsv loses it.
The 25–61 species left are still well above the 9–16 genomes sylph finds. Compare otu_table.tsv with the gate off and on before relying on auto, and do not take the estimate of dropped read pairs in logs/kmer_evidence/{sample}.log for the measured change: with auto it was 3.5 and 3.9 times too low in the two samples that lose a whole order, because it only sees the reads Kraken2 assigned, not those Bracken moves.

Why the default is `enabled: false`¶

The gate is off by default because a correctly scaled Bracken threshold already removes most of what it would remove. It has been run on 46 samples: the D6331 mock, the four PRJNA1477390 metagenomes and 41 CAMI II samples, and on the existing Kraken2 reports of the MBARC-26 mock and seven forest soils (above). In gate-on/gate-off runs of CAMI II samples at several depths with PlusPF k2_pluspf_20251015, the gate changed species-level F1 by a mean of +0.0009 at the depth-scaled threshold, and by zero on the strain-madness and rhizosphere samples at every depth tested. At the former fixed -t 10 the mean was +0.0079, about nine times more. So the gate mainly helps when the Bracken threshold is not scaled to depth; with a scaled threshold it adds little.

Which database build those two numbers come from

Both figures come from gate-on/gate-off runs on PlusPF k2_pluspf_20251015 (15 sample-depth combinations at the scaled threshold, 11 at the fixed one). The k2_pluspf_20260626 benchmark runs had the gate off; removing from their Bracken output the species the gate would prune at the default floor of 300 changes mean species-level F1 by +0.0003.

Unchanged F1 does not mean an unchanged table¶

The two filters ask different questions. The threshold asks how many reads; the gate asks how much distinct sequence. A taxon can pass one and fail the other, and an aggregate score of 250+ species barely notices when a single one flips.

Pseudomonas sp. JS425 in the four test samples, with the Bracken threshold scaled to depth (threshold_alpha 5×10⁻⁵):

Sample	Classified pairs	Bracken `-t`	Kraken2 pairs	Distinct minimizers	Passes `-t`	After Bracken, gate off	Gate
SRR39106521	2,268,774	113	372	182	yes	69,992	removed
SRR39106522	3,458,847	173	103	181	no	–	removed
SRR39106523	1,673,810	84	301	181	yes	51,418	removed
SRR39106524	2,177,429	109	123	181	yes	8,935	removed

In SRR39106521 the 372 read pairs pass the threshold, but they rest on 182 distinct minimizers, each hit about 60 times, and Bracken raises the call to 3 % of the sample. The threshold keeps JS425 in three samples and drops it from the fourth only because the cut-off falls differently there; in a presence/absence plot that would look like a real difference between samples. The gate removes it from all four.

These samples have no known composition, so they do not enter the F1 scores above. That is why the benchmark shows almost no gain from the gate while the table changes.

When to consider turning it on: communities the database covers poorly, where reads from organisms absent from the database are assigned to their relatives. The evidence for this is limited. It rests on cases like JS425 in the wastewater test samples, which have no known composition; on the CAMI II rhizosphere samples, the poorly covered benchmark set with a known composition, the gate hardly changed F1. For gut, clinical or other samples close to RefSeq, leave it off. Either way the gated report is written even when the gate is off, so compare it with the Kraken2 report before deciding.

Validation status¶

What the gate was tested against, and what those tests could and could not show.

What has been checked¶

The gate was checked on one real mock community (Zymo D6331, below) and on CAMI II's marine, rhizosphere and strain-madness gold-standard sets (41 samples): in gate-on/gate-off runs with PlusPF k2_pluspf_20251015, and by applying it to the Bracken calls of the k2_pluspf_20260626 runs. These results were scored with the benchmark's own scorer. On the k2_pluspf_20260626 runs (Kraken2 confidence 0.15 and 0.20) that scorer was checked against OPAL 1.0.13, the CAMI assessment tool: given the same input preparation, OPAL's species-level counts of true positives, false positives and false negatives equal the scorer's in every sample. The full methodology and the numbers that matter for choosing a threshold are on the confidence and threshold page. The result for this gate: once the Bracken threshold is scaled to depth it adds at most 0.002 F1 (PlusPF k2_pluspf_20251015), but it still removes calls like JS425 above, which a read-count threshold cannot detect. The second mock on this page, MBARC-26, and the seven forest soils were used only to test the auto floor beyond the benchmark's depth (above); they do not enter these scores.

One real mock community: the ZymoBIOMICS Gut Microbiome Standard (D6331), SRA run SRR29240428 (BioProject PRJNA1118346, Martí et al. 2025; see Benchmark data), randomly subsampled so that 5.30 million read pairs remained after fastp, of which Kraken2 with PlusPF k2_pluspf_20251015 classified 5.29 million.

The standard is sold as 21 strains. A species-level profiler can only ever see 17 species, because five of the 21 are Escherichia coli strains at 2.8% each that collapse into one E. coli row at 14%. That ceiling is arithmetic, not a reference-database gap. Zymo lists the standard's Fusobacterium as F. nucleatum (strain 2/1/50A), but sylph (GTDB r232), MetaPhlAn 4 and Kraken2 with GTDB r226 or PlusPF k2_pluspf_20260626 place most of these reads in F. animalis, a former subspecies of F. nucleatum now ranked as a species. It is therefore scored as F. animalis, and calls of F. nucleatum, F. polymorphum and F. vincentii count as false positives. Homo sapiens, from human DNA in the sample, is not scored.

The run was processed through the pipeline with PlusPF k2_pluspf_20251015, Kraken2 confidence 0.15, Bracken at the former fixed -t 10 and the gate off. For the gate-on column, Bracken was run again with the same settings on the gated report that the pipeline writes at a floor of 333, the file it hands to Bracken when the gate is on. A floor of 300 gives the same species list on this run. At the current default threshold (-t 264 for this run), only 2 of the 29 false positives in the table below have enough reads to pass (F. nucleatum and F. polymorphum), as does the unscored Homo sapiens; the gate removes none of them, and the threshold itself drops Salmonella enterica.

	Gate off	Gate on
True members recovered	15 / 17	15 / 17
False-positive species	29	18
Precision	0.341	0.455
F1	0.492	0.600

No true member was removed. Salmonella enterica, 0.01% of the standard and 71 classified read pairs, has 3,400 distinct minimizers at a duplicity of 1.2, more than ten times the floor, and is kept. The gate depends on this difference: a rare species can still have many distinct minimizers.
The two members never recovered (Enterococcus faecalis at 0.001%, Clostridium perfringens at 0.0001%) have 2 and 5 Kraken2 read pairs, below Bracken's -t 10, so neither reaches the table, with the gate off or on. Both have species entries in the database, so they were missed for lack of reads.
Quantitative accuracy did not change. With the human row left out, L1 distance to the theoretical composition was 27.80 without the gate and 27.79 with it, Shannon H 2.321 and 2.320, and the Bray–Curtis dissimilarity between the two profiles 0.0001. The 43 taxa the gate removed at 333 held 445 of the 5.3 million read pairs (the same 43 taxa at 300). Any ordination or differential-abundance test would give the same answer from both tables.
Richness, on the other hand, improved substantially — which is where the gate did its work on this dataset:

Observed richness	Truth	Gate off	Gate on
Species	17	44 (+159%)	33 (+94%)
Genera	17	24 (+41%)	16 (−6%)

On this mock the gate changed which species are listed, not their abundances. If you report how many taxa you detected, or work with presence/absence, that matters. If you only ever compare relative abundances between samples, on data like this it will not change your conclusions.

Why richness is still over-estimated with the gate on¶

With the gate on, 33 species other than human are still reported against 17 in the standard. Most of the 18 false positives are not the kind of call the gate is designed to remove.

Of the 18 false-positive species that survive, 17 are congeners of a genuine member — and the other one is defensible. The table also lists Homo sapiens, which is not scored:

Survivor	Reads after Bracken	Distinct minimizers	What it is
Fusobacterium nucleatum, F. vincentii, F. polymorphum	226,289	1,113–30,392	Close relatives of the true member, F. animalis; all four were subspecies of F. nucleatum until F. animalis, F. vincentii and F. polymorphum were raised to species rank. With `k2_pluspf_20251015`, Kraken2 gives F. nucleatum more reads than F. animalis (18,038 against 7,570)
Homo sapiens	1,802	15,682	Trace human DNA from handling; present in the sample, a contaminant that is not scored
Staphylococcus epidermidis	22	1,212	A common skin bacterium and contaminant; plausibly present in the sample
14 further congeners of Veillonella, Fusobacterium, Saccharomyces, Bacteroides, Candida, Clostridioides	1,543	566–10,344	Close relatives of an abundant true member

All these survivors carry large distinct-minimizer counts — Saccharomyces paradoxus has 10,344 on just 144 read pairs, at a duplicity of 2.0. That is broad, un-repetitive evidence, and it is genuinely in the read data: S. cerevisiae is a real member of the standard, and its sequence matches a great deal of what the database files under its sister species.

So there are two different ways a species-level call goes wrong, and only one is an evidence problem:

Thin evidence — a few conserved minimizers hit over and over. Low distinct count, high duplicity. The gate removes these.
Mis-attributed evidence — a genuinely present organism whose sequence matches a close relative's database entry. High distinct count, low duplicity. The gate does not remove these, and should not: the sequence is in the sample, and only the species label is wrong.

The residual inflation is almost entirely the second kind, which is a property of how finely the reference database splits closely related organisms — not something any k-mer counting method can detect. The practical consequence is visible in the table above: at genus rank, where those distinctions disappear, the gate brings richness from 41% above the true value to 6% below it: the 16 genera left are 15 of the 17 true ones (the two members that never reach the table are missing) and Staphylococcus.

What this does not show¶

On D6331 the gate removed no true species; on CAMI II it changed F1 very little

Every organism in D6331 has a species-level entry in the database used. When that is true, classification is already close to correct, most false positives are a thin tail of taxa with few reads, and there is little for an evidence filter to repair, which is what the CAMI II benchmark found: with the Bracken threshold scaled to depth, this gate adds at most 0.002 F1 across marine, rhizosphere and strain-madness samples (PlusPF k2_pluspf_20251015).

The gate is still useful in cases like JS425 above: a taxon with an ordinary read count but few distinct minimizers. A read-count threshold cannot detect this, however it is scaled.

The CAMI II benchmark described on the confidence and threshold page covers what one mock cannot: larger and more even communities (marine, 256–381 species per sample), several samples at several depths, communities with genomes absent from the reference (rhizosphere: only 165 of its 307 gold-standard species, 54 %, are in PlusPF k2_pluspf_20260626), and sample types beyond gut and wastewater. It agrees with the D6331 result that the gate removes no true species from the final table at full sequencing depth (on MBARC-26 it removed one thin member from the Kraken2 report, which Bracken's read threshold drops anyway), and it shows that once the Bracken threshold is scaled to depth the gate changes F1 very little.

Still untested: everything above used short-read Illumina data. Long-read or hybrid data, and sample types with sparser reference coverage than rhizosphere, have not been tried. If you switch the gate on, check the evidence tables for what it removed.

Limitation: very close relatives¶

Very close relatives defeat the method, and no k-mer evidence measure fixes this. Shigella dysenteriae is not in D6331, and only 3 read pairs were assigned to it, too few to pass Bracken's -t 10. Yet the sample holds 1,042 distinct minimizers that the database marks as specific to it, more than three read pairs can carry and more than three times the floor, so the gate would keep it. Shigella and E. coli are similar enough that minimizers the database marks as Shigella-specific are genuinely present in E. coli reads. That is a property of the reference taxonomy, not something this filter can detect.

Threshold stability¶

A fixed threshold should give a species the same verdict in every sample, but it does not always. Of 620 species found in more than one of the four PRJNA1477390 test samples, 123 (20 %) got different verdicts at the floor of 300:

62 % differ mainly in depth: the species had at least five times fewer reads in one sample, and with few reads it cannot reach enough distinct minimizers. This change of verdict is probably correct.
38 % have read counts less than five-fold apart. 20 of these 47 species lie close to the floor, between 200 and 450 distinct minimizers, in every sample. Sphingomonas paucimobilis is one: 155 to 199 read pairs in each sample, 350 distinct minimizers in SRR39106524 (kept) and 277 to 295 in the other three (removed). The other 27 lie further from the floor in at least one sample, sometimes much further: Sphingopyxis lindanitolerans has 93 to 105 distinct minimizers in three samples and 10,626 in the fourth, on 95 to 282 read pairs.

Any fixed cut-off behaves like this, whatever its value. Check the evidence tables for species whose verdict differs between samples.

What does not work¶

Three approaches were tested on the four PRJNA1477390 test samples and rejected. They are described here because each seems reasonable before it is measured.

A threshold on distinct minimizers alone is partly an abundance filter. The distinct count correlates with read count (Pearson r = +0.71 on log-log across the 2,418 species rows of the four Kraken2 reports), so some of what it removes, it removes for being rare. This is why the published thresholds pair the k-mer count with an explicit read floor — min_reads exists for exactly that, and is 0 by default only because Oskolkov (2026) found that filtering on unique k-mers alone was enough to reach the best F1, with no read-count filter needed.

Normalising to minimizers-per-read inverts the filter. That ratio falls as abundance rises (r = −0.72): across the four samples its median is 84 for single-read species and 1.4 for species with 1,000 reads or more. A "ratio below 5" rule on SRR39106521 removed 80 % of the reads that had reached species rank, including the two species with the most reads, while keeping all 160 single-read species.

Thresholding duplicity removes the most abundant species. Across all species rows duplicity hardly changes with abundance (r = +0.17), but the most deeply sequenced species have high values, because each of their minimizers is seen many times. Achromobacter denitrificans, the largest species in all four samples, has a duplicity of 23 to 74 on 42,000 to 1.1 million distinct minimizers. A duplicity cut-off of 10 prunes about as many species rows as the floor of 300 (65 % against 66 %), and its verdicts differ between samples less often (14 % of species against 20 %). But it would remove 73 to 93 % of each sample's table, A. denitrificans included, where the floor of 300 removes 1.3 to 5.8 %. With the gate's 1 % safety rail, which would keep A. denitrificans, it would still remove 16 to 65 %. Duplicity is therefore reported, not thresholded.

Flagging outliers against a fitted trend line finds no distinct population. Fitting log(distinct minimizers) against log(reads) and taking residuals looks principled, but the residuals are essentially normal: in SRR39106521 a "2 SD below the line" rule flagged 14 of 578 species, where a normal distribution predicts 13.1. There is no separable cluster to find, so the rule just removes the bottom few percent of any sample. Worse, because the line is fitted per sample, the same species can fall on opposite sides of it in different samples of one study.

The conclusion for the design: judge every species against a floor that is not fitted to the evidence of the other species in its sample, and treat the trend line as a ranking for the eye rather than a test. The floor is one fixed number for every sample of a run or, with auto, a number set by the read pairs Kraken2 classified in the sample, human ones left out. With auto the same species can therefore get different verdicts in samples of different depth. That is intended when the whole community was sequenced deeper: a species that is really present then leaves more distinct minimizers (see Scaling the floor with depth). But the floor follows the depth of the whole sample, not the reads of each species, so a species with the same reads in a deeper sample, where it is a smaller share, faces a higher floor. In the test dataset Novosphingobium pentaromativorans has 2,156 to 2,195 distinct minimizers in all four samples, on 2,955 to 3,550 read pairs from Kraken2; auto keeps it in the two samples with floors of 1,673 and 2,176 and removes it from the two with floors of 2,267 and 3,458. In the deepest of the four, 57 % of the classified pairs belong to one species, Achromobacter denitrificans, which raises the floor for everything else without adding to their evidence.