Databases

Databases

Any Mash sketch database works with mashID. Databases are stored outside the package and outside the repository.

Pre-built databases

mashID_download_db                    # list databases and whether they are installed
mashID_download_db mycobacteriaceae   # download one (MD5-verified)
mashID_download_db --all --dir /db    # download everything somewhere else

The databases dated 2026-10-22 need mashID 0.3.0 or newer, and the 2026-10-23 brucella rebuild needs 0.3.1: the download registry ships in the package, so an older install silently fetches an older set (mashID --version to check, and mashID_download_db with no arguments to list what your copy knows about).

Sizes are the whole download: sketch plus any sidecars. mashID_download_db with no arguments lists the sketch alone, and a fetch prints each file's size as it goes, so neither prints this total. Mebibytes, as everywhere else in the output.

Name	Content	Size
mycobacteriaceae (default)	Mycobacteriaceae (NCBI taxon 1762): 16,433 GenBank assemblies as of 2026-10-22, dereplicated per species at 99.9% identity to 3728 references, NCBI TaxIDs for all. k=21, s=10000. Ships with its metadata and confusable sidecars. Figshare	286.9 MB
mycobacteriaceae-2025	The previous Mycobacteriaceae database (2025-02-20): 3309 references, k=21, s=1000, no TaxIDs. Kept for reproducibility. Figshare	26.4 MB
listeria	Listeria (NCBI taxon 1637): 7453 RefSeq assemblies as of 2026-09-22, dereplicated per species at 99.9% identity to 1441 references covering 42 species and subspecies, NCBI TaxIDs for all. k=21, s=10000. Ships with its metadata sidecar. Figshare	110.8 MB
listeria-2025	The previous Listeria database (2025-02-18), no TaxIDs. Kept for reproducibility. Figshare	46.3 MB
salmonella	Salmonella (NCBI taxon 590): 21,543 RefSeq assemblies as of 2026-10-22, binned by subspecies and dereplicated to 1436 references covering S. bongori and all six S. enterica subspecies, NCBI TaxIDs for all, and 29 mislabelled references corrected before binning. k=21, s=10000. Ships with its metadata and confusable sidecars. Species and subspecies level: the serovar in a reference name is that of the nearest reference, not a serotype determination; use SeqSero2 or SISTR for serotyping. Figshare	110.5 MB
brucella	Brucella (NCBI taxon 234, including the former Ochrobactrum species): all 2022 GenBank assemblies as of 2026-10-23, dereplicated per species at Mash distance 0.0003 to 389 references covering 29 species, NCBI TaxIDs for all, and 13 mislabelled references corrected before binning. k=21, s=10000. Ships with its metadata and confusable sidecars. Held-out concordance 96.8%, and every miss carries a "not separable" note; the classical species are host-adapted lineages 0.2 to 0.5% apart, so their calls are nearest-reference calls and biovars are not resolved. Figshare	30.0 MB
escherichia	Escherichia and Shigella (NCBI taxa 561 and 620): 58,198 RefSeq assemblies as of 2026-10-22, dereplicated to 4510 references (five other Escherichia species and four Shigella species beside E. coli), NCBI TaxIDs for all, and two mislabelled Shigella references corrected before binning. k=21, s=10000. Ships with its metadata and confusable sidecars. Held-out concordance 97.1% at species level on the 2026-10-22 build; S. boydii, S. dysenteriae and S. flexneri share lineages and are not reliably separated from each other, nor enteroinvasive E. coli from Shigella: use ShigaTyper or ShigEiFinder. Figshare	346.6 MB
refseq_viral	All 19,625 viral genomes and genome segments of NCBI RefSeq release 237 (6427 phages), current ICTV species names and TaxIDs from the GenBank flat files. k=21, s=2000. Ships with its metadata and confusable sidecars. Median reference 6.9 kb, so mashID applies no reference-length filter. For virus identification and for prophage or viral contamination detection in any sample. Figshare	294.1 MB
progenomes4	proGenomes 4 representative genomes: 32,887 species-cluster representatives across 4384 genera of bacteria and archaea, current NCBI names and TaxIDs for all. k=21, s=2000. Ships with its metadata sidecar. For species-level screening of anything. Figshare	519.3 MB
progenomes3	proGenomes v3 representative genomes (2023), names parsed from headers. Superseded by progenomes4. Figshare	330.9 MB
refseq_bacteria	RefSeq bacteria as of 2023-01-19, dereplicated at 99% identity: 82,557 references across 32,750 organisms. It also ships a metadata sidecar, added on 2026-10-22, carrying current NCBI names and TaxIDs for 82,547 of them, so its names no longer come from the reference headers. k=21, s=1000. Three years old and superseded in practice by progenomes4, which covers the same broad-screening niche with current names; kept for reproducibility and ships no confusable table. Figshare	671.2 MB

Databases live in $MASHID_DB_DIR if set, else $XDG_DATA_HOME/mashID/db (~/.local/share/mashID/db). Pass a name (-d listeria) or a path (-d /db/x.msh) to mashID. A metadata sidecar (see below) is generated right after download.

Building your own

make_mashID_db -i /path/to/genomes -o /path/to/db -p my_database -s 10000 -k 21 -t 16
# -> /path/to/db/my_database.msh and /path/to/db/my_database.metadata.tsv

Option	Description
-i PATH	Directory searched recursively for .fna/.fa/.fasta[.gz], or a text file listing one genome path per line.
-o DIR, -p NAME	Output directory and database name (.msh is appended).
-s 10000	Sketch size. 10000 is recommended for species-level identification.
-k 21	K-mer size (1–32).
--min-length 100000	Exclude references shorter than this; use 0 for plasmid or viral collections.
--metadata FILE	TSV/CSV with Accession, Organism and optionally TaxID columns.
--assembly-report FILE.jsonl	assembly_data_report.jsonl from an NCBI datasets download genome archive.
--annotate DB.msh	Only (re)write the metadata sidecar of an existing database.
--check DB.msh	Report quality issues of a database: references shorter than --min-length, unparsed organism names, missing TaxIDs, species listed under several genera (synonyms), single-reference organisms.

scripts/build_taxon_db.sh in the repository shows a full workflow: download all NCBI genomes of a taxon with datasets, dereplicate them per species with Assembly-dereplicator, and sketch the database.

Validating a database on held-out genomes

scripts/validate_db_heldout.py <build_dir> <db.msh> screens genomes that the build downloaded but did not keep (dropped by the bin cap or by dereplication) and reports, per species, how often mashID returns the NCBI name, how often a note fired, and every mismatch. Run it after each build; the numbers in the provenance table above come from it.

Checking a database

make_mashID_db --check /path/to/db.msh

Run this on any database before trusting it. On the 2025 Mycobacteriaceae database (mycobacteriaceae-2025) it reports seven references under 100 kb (partial records that would otherwise become false top hits), no TaxIDs, and 40 species that appear under two or three genera because of the Mycobacterium → Mycolicibacterium / Mycobacteroides / Mycolicibacter reclassification. The last point matters: mashID counts Mycobacterium abscessus and Mycobacteroides abscessus as different organisms, so a sample may show a spurious "Possible mixture" between the two names. A sidecar built from NCBI's assembly report (--annotate --assembly-report) or a curated --metadata table harmonises them.

Provenance and known issues of the pre-built databases

Database	Built	Source and filters	Known issues
mycobacteriaceae	2026-10-22	scripts/build_taxon_db.sh with CORRECTIONS=curation/label_corrections.tsv: all 16,433 GenBank assemblies of taxon 1762 (atypical excluded) via NCBI Datasets, those corrections applied to the binning and to the sidecar, binned by species from the assembly report, dereplicated per species at Mash distance 0.001 with Assembly-dereplicator 0.3.2, references < 100 kb excluded, sketched k=21, s=10000. 3728 references; names and TaxIDs from the assembly report, strain text removed.	--check on this build reports 3728 references, TaxIDs for all, 640 organism names, and one species epithet under two genus spellings (NCBI's own naming). Held-out validation on 623 genomes from 155 species groups: 98.4%. Of the 10 misses, three are Mycobacteroides/Mycobacterium chelonae, that same genus synonym; four are M. pseudoshottsii matching M. marinum, which the confusable table flags; two are the corrected references below, where the call matches the corrected name and the deposited one is what counts as the miss; one M. tuberculosis matched an "sp." reference. MTBC members are not separable (see Interpreting results). Two references were mislabelled at NCBI (GCA_013697125.1, deposited as "Mycobacterium canetti", is M. tuberculosis variant bovis; GCA_025822535.1, "Mycolicibacterium goodii", is M. wolinskyi; see curation/label_corrections.tsv). This build applies both before binning, and neither genome survives: filed with its true relatives each was a near-duplicate and dereplication dropped it. Databases downloaded before 2026-10-22 carry the deposited names; re-run mashID_download_db mycobacteriaceae.
mycobacteriaceae-2025	2025-02-20	All NCBI assemblies of taxon 1762 from the datasets web table (26,101), renamed and binned by the species in the first header, dereplicated per species at Mash distance 0.001 with Assembly-dereplicator 0.3.2, sketched k=21, s=1000. 3309 references.	Seven references < 100 kb; names parsed from headers, so genus synonyms are not merged; no TaxIDs; s=1000 gives identities in steps of 0.001. Superseded by mycobacteriaceae.
listeria	2026-09-22	scripts/build_taxon_db.sh with ASSEMBLY_SOURCE=RefSeq and taxon 1637: all 7453 RefSeq assemblies (atypical excluded; GenBank's 79,000 are mostly redundant L. monocytogenes), binned by species from the assembly report, dereplicated per species at Mash distance 0.001 (L. monocytogenes: 892 kept), references < 100 kb excluded, sketched k=21, s=10000. 1441 references with NCBI names and TaxIDs.	--check reports no issues. Validated on Nanopore L. ivanovii subsp. londoniensis and L. monocytogenes reads and an assembly.
listeria-2025	2025-02-18	All NCBI Listeria assemblies, dereplicated as above.	Names parsed from headers; no TaxIDs. Superseded by listeria.
salmonella	2026-10-22	scripts/build_taxon_db.sh with ASSEMBLY_SOURCE=RefSeq BIN_RANK=subspecies MAX_BIN=3000 CORRECTIONS=curation/label_corrections.tsv and taxon 590: all 21,543 RefSeq assemblies (GenBank's 630,000 are mostly redundant), 29 curated label corrections applied, binned by subspecies, each bin capped at 3000 keeping complete genomes first, dereplicated per bin at Mash distance 0.001, assemblies named only "Salmonella sp." excluded, sketched k=21, s=10000. 1436 references with NCBI names and TaxIDs.	Nearest-reference serovars appear in names but Mash does not serotype: serovars are defined by three loci, several are polyphyletic, and rare ones have few references. --check reports no issues on this build: 1436 references, TaxIDs for all, 175 organism names. Validated on SKESA assemblies and MiSeq reads of S. enterica. Held-out validation of 107 genomes at subspecies rank places none of them in the wrong subspecies, but the raw concordance is 68/107 because the deposited name is the yardstick: 23 of the held-out genomes are ones this project has corrected, so a right call counts as a miss, and the other 16 differ only in granularity — 9 matched a reference whose NCBI name carries no subspecies, and 7 are bare "Salmonella enterica" genomes called to a subsp. enterica serovar. Read the figure as a measure of agreement with NCBI labels, not of accuracy.
brucella	2026-10-23	scripts/build_taxon_db.sh with ASSEMBLY_SOURCE=GenBank EXCLUDE_SP=1 DEREP_DISTANCE=0.0003 CORRECTIONS=curation/label_corrections.tsv and taxon 234: all 2022 GenBank assemblies, 13 curated label corrections applied, binned by species, dereplicated per species at Mash distance 0.0003 (0.001 left only 21 B. melitensis and 9 B. suis references and scored 90.7% held-out; the classical species are only 0.2 to 0.5% apart), "sp." bin excluded, sketched k=21, s=10000. 389 references.	--check reports no issues. Held-out validation of this build on 185 genomes from 24 species groups: 96.8%, up from 93.0% before the ten corrections of 2026-10-23, with B. abortus, B. canis, B. melitensis and B. suis each 20/20. The gain was predicted and then measured: seven misses had been held-out B. suis genomes matching one reference deposited as B. melitensis, which the published species-specific qPCR assays, MLST-9 and this validation all identify as B. suis. All six remaining misses are the submission batch deposited as B. intermedia that is identical to B. ciceri, and the confusable table is down from nine sub-threshold cross-name pairs to two — that batch, and the canis/suis boundary, which is not a labelling error but a lineage inside a species. For genera whose species are this close, set DEREP_DISTANCE below the inter-species distance.
escherichia	2026-10-22	scripts/build_taxon_db.sh with ASSEMBLY_SOURCE=RefSeq MAX_BIN=5000 EXCLUDE_SP=1 CORRECTIONS=curation/label_corrections.tsv and taxa 561,620: all 58,198 RefSeq Escherichia and Shigella assemblies, 2 curated label corrections applied, binned by species, each bin capped at 5000 keeping complete genomes first, dereplicated per bin at Mash distance 0.001, "sp." bins excluded, sketched k=21, s=10000. 4510 references.	--check reports no issues: 4510 references, TaxIDs for all. Validated with scripts/validate_db_heldout.py on 171 genomes left out of this build: 97.1% species concordance, and all five misses are Shigella genomes matching another Shigella species, each carrying a "not separable" note — a labelling matter Mash cannot resolve. Complete genomes over-represent clinical and outbreak lineages.
refseq_viral	2026-09-28	scripts/build_virus_db.sh: the viral division of RefSeq release 237 from the NCBI FTP release files, split into one file per record, names and TaxIDs from the GenBank flat file (scripts/refseq_gbff_metadata.py), sketched k=21, s=2000, no length filter, no dereplication (one reference per species or segment). 19,625 references.	Names are RefSeq's ORGANISM field, i.e. ICTV species names ("limnipivirus B1"); the familiar name is in Description. Segmented viruses appear as separate segments. Divergent viruses without a close RefSeq relative are not detected at k=21. Tested: RefSeq phages self-match at identity 1; two S. Infantis assemblies show a complete Salmonella phage MET_P1_103_31 prophage at 0.9996; M. bovis assemblies and reads give no hit.
progenomes4	2026-09-22	scripts/build_progenomes_db.sh: the proGenomes 4 representatives multi-fasta (45.9 GB) split per assembly accession, names and TaxIDs fetched from NCBI Datasets for every accession, sketched k=21, s=2000, no dereplication (one genome per species cluster already). Its Source_File column names that archive (https://progenomes.embl.de/data/repGenomes/pg4_genomes_representatives.fna.gz) rather than the per-genome file the build split out of it: the split directory was a 45.9 GB intermediate and was deleted after the sketch, and each file was named <Accession>.fna, so the Accession column already says which record of the archive a reference is.	Single representative per species, so identities of 0.95 to 0.99 are normal; use a group-specific database for finer resolution. --check reports 2230 epithets shared across genera and 91 references named "bacterium" or similar: expected at this scale, not synonyms.
progenomes3	2023-03	proGenomes v3 representative genomes, sketched from the proGenomes fasta.	Header format differs from NCBI's; names come from the sidecar generated on first use, check them with --check. Superseded by progenomes4.
refseq_bacteria	2023-01-19	RefSeq bacteria, dereplicated at 0.01. A metadata sidecar was added on 2026-10-22 from NCBI Datasets (make_mashID_db --annotate --assembly-report), without touching the 2023 sketch.	Large (671 MB with the sidecar); three years old.

mycobacteriaceae-2025, listeria-2025 and progenomes3 ship without a metadata sidecar, so mashID parses each reference's name from its own fasta header on every run. Two consequences follow. First, a change to the name parser reaches them with no file to update, which is how the fix that stopped truncating Salmonella antigenic formulae at the first comma took effect for them. Second, a curated label correction cannot be applied to them at all: scripts/apply_label_corrections.py keys on an assembly accession and their sketch IDs carry none, being named after a contig (JACEGV010000031.1_Mycobacterium_canetti_strain_ITM_C02321) or a proGenomes sample, which is also what the Accession column reports for them. So mycobacteriaceae-2025 still identifies the mislabelled GCA_013697125.1 as M. canetti, and only mycobacteriaceae carries that correction.

refseq_bacteria was in that group until 2026-10-22. Every one of its sketch IDs does carry a GCF accession, so it could be annotated without a rebuild, and since v3 it ships a sidecar: names and TaxIDs from NCBI Datasets for 82,547 of its 82,557 references, the other ten being RefSeq records withdrawn since 2023, which keep their header names and have no TaxID. The five curated corrections that fall inside it are applied, one of them only because a correction now matches either accession of an assembly — the curation file records that genome as GCA_025822535.1 while this database carries GCF_025822535.1. The 2023 sketch itself is unchanged, so mashID_download_db refseq_bacteria fetches only the new sidecar.

scripts/build_virus_db.sh builds a database of every RefSeq viral genome from the NCBI RefSeq release files, with names and TaxIDs from the GenBank flat files (scripts/refseq_gbff_metadata.py).

The metadata sidecar

<database>.metadata.tsv sits next to the .msh file and holds, for every reference: Accession, Organism, TaxID, Length, Hashes, Source_File, Description.

It is the source of Identification and TaxID in the outputs, and of the reference lengths used by --min-ref-length. Organism names come from, in order of preference:

1. --metadata FILE: your own table.
2. --assembly-report: NCBI's report, which carries organism names and TaxIDs for every accession.
3. The first fasta header of each genome, parsed for genus, species and infraspecific ranks (NZ_CP060409.1 Mycolicibacterium fortuitum strain W4 chromosome → Mycolicibacterium fortuitum).

Accessions are recognised in file names (GCF_000195955.2_ASM19595v2_genomic.fna.gz → GCF_000195955.2). If a database has no sidecar, mashID builds one on first use from mash info, and writes it next to the database when the directory is writable. To add or refresh a sidecar, including for downloaded databases:

make_mashID_db --annotate /path/to/db.msh --assembly-report assembly_data_report.jsonl

Organisms the database cannot separate

Whether two organisms can be told apart is a property of the database, not of a sample: the same two references are the same distance apart in every run. Computing it once makes the per-sample annotation free, and correct whether or not winner-take-all is in use — unlike --ambiguity-margin, which cannot see a close relative under winner-take-all at all (see Outputs).

make_mashID_db --confusable /path/to/db.msh -t 16          # writes db.confusable.tsv
make_mashID_db --confusable /path/to/db.msh --confusable-distance 0.002

For each reference the table records the closest reference of every other organism within --confusable-distance (default 0.005, matching --ambiguity-margin), nearest pair first:

Accession        Organism                                  Confusable_Organism             Distance  Nearest_Accession
GCA_025821985.1  Mycobacterium timonense                   Mycobacterium bouchedurhonense  0.000041  GCA_025822125.1
GCA_013697125.1  Mycobacterium tuberculosis variant bovis  Mycobacterium tuberculosis      0.000972  GCA_021325765.1

mashID picks the sidecar up automatically and notes affected calls; --db-confusable points at a different table. A pre-built database that publishes the table alongside it installs the table too, so the O(n^2) computation only has to happen once, for whoever built the database. A database downloaded before its table was published picks the table up on the next mashID_download_db <NAME> for it, fetching only the sidecars and leaving the sketch alone — the listing calls a database installed on the strength of its .msh, so it will say yes while the sidecars are still missing. Runs that finished before the table existed do not have to be screened again either:

scripts/reannotate_summary.py /path/to/results --dry-run   # then without --dry-run

rebuilds their summary files from the per-sample tables and the current sidecars, which is also how to apply a corrected organism name to results that predate the correction.

Five of the pre-built databases ship a table: mycobacteriaceae, salmonella, brucella, escherichia and refseq_viral. Two were checked on 2026-10-23 and need none, which is worth stating rather than leaving silent, since a missing table otherwise looks like an omission:

database	references within 0.005 of each other	pairs of different organisms
listeria	11,660 pairs	none: every species in it is separable
progenomes4	no pair at all	none, necessarily: it holds one representative per species cluster, so nothing is that close to anything else
refseq_viral	—	765 rows over 421 of its 19,625 references

Most of the refseq_viral rows are what the table is for: distinct entries that k=21 cannot separate, such as the Flavobacterium phages Fpv5, Fpv8, Fpv10 and Fpv11 at 0.000012 from one another. Forty rows are a different thing worth knowing about — one virus under two names, at distance zero, where RefSeq carries both the ICTV name and the older familiar one (Canine adenovirus 1 and Canine mastadenovirus A, Human adenovirus 2 and Human mastadenovirus C). The note is still true there, but it means "either name may be reported", not "two candidate organisms".

refseq_bacteria is deliberately not checked and will not be: 82,557 references is some 3.4 billion pairs, for a 2023 sketch whose niche progenomes4 now covers with current names.

The contrast with brucella is the point of the whole table — there, every held-out misidentification carries a "not separable" note, because its species are one genomospecies.

Which names count as different organisms

Names are compared as taxonomic claims, not as strings. A name is read as a chain of ranks, and a pair is reported only when the two chains genuinely disagree:

Pair	Reported	Why
Mycolicibacterium goodii / M. wolinskyi	yes	two species
M. tuberculosis / M. tuberculosis variant bovis	yes	NCBI uses variant where it used to use species: M. bovis, M. africanum, M. microti and M. caprae are all "variants" now, and they are what a reader still means by different organisms
Mycobacteroides abscessus subsp. abscessus / subsp. massiliense	yes	two subspecies of one species disagree with each other
Mycobacterium avium / M. avium subsp. hominissuis	no	one is the other at a finer rank; a sample that is the subspecies is also, correctly, the species
Mycobacterium abscessus / Mycobacteroides abscessus	no	a renamed genus, not a second organism
Escherichia coli / E. coli O157:H7	no	a strain designation below the species
Mycobacterium fortuitum / Mycobacterium sp.	no	Genus sp. claims no species, so it contradicts none
[Mycobacterium] chelonae / Mycobacterium chelonae	no	brackets, Candidatus and letter case are notation, not taxonomy

Letting a species pair with its own subspecies was two fifths of the rows of the mycobacteriaceae table (767 of 1,993) and the two commonest notes the database produced — "M. avium subsp. hominissuis not separable from M. avium" and the abscessus equivalent — neither of which tells anyone anything. Reading the ranks instead leaves 1,226 rows over 313 references rather than 945, and keeps every pair that matters, the whole M. tuberculosis complex among them.

The suppression is symmetric: a call of M. avium is not told it could be M. avium subsp. avium either. That direction is a statement about resolution rather than about a competing organism — mashID calls species, and a reference named only to species cannot place a sample below it — and keeping it would have left most of the noise in place, since the plain species name is the one most references carry.

The rule is applied both when the table is written and when a note is built from it, so a table produced by an earlier version, or by hand, cannot reintroduce those rows.

These are still names being compared, not a taxonomy, and two cases cannot be settled from the strings alone:

• The same epithet under two genera is read as a rename. For bacteria within --confusable-distance that is almost always right — two genera really 99.5% identical are one renamed taxon — but it is an assumption, and it is wrong wherever the first word is not a genus. In refseq_viral, where the first word is a host or a place, Aeromonas phage D3 and Pseudomonas phage D3, or Wuhan and Xingshan cricket virus, are read as one organism. It also blinds the mislabel use below to a genome deposited under a homonym (Camelimonas abortus for Brucella abortus).
• One taxon under NCBI's old and new spelling — Mycobacterium bovis and M. tuberculosis variant bovis — is reported as a conflict. The shipped mycobacteriaceae sidecar uses one spelling throughout, but a database built from an older NCBI dump, or a hand-written correction, can mix them. make_mashID_db --check lists such pairs; harmonise them with --metadata.

Cost, and reading the table for mislabels

It is an all-pairs comparison, so cost grows with the square of the reference count: 3,534 references of mycobacteriaceae take under two minutes and 850 MB, but a database with hundreds of thousands of references is a different proposition. This is why it is a separate step rather than part of every build.

Read the table top-down and it doubles as the sharpest mislabel detector available: a reference named for one species sitting a hair from another species' references is either a genuinely inseparable pair or a bad label. Both mislabelled mycobacteriaceae references were in the first 20 of the 2,102 rows the table had before they were corrected, GCA_025822535.1 at rank 4.

Correcting a mislabelled reference

Some assemblies are deposited under the wrong organism name, and a reference named for a species it does not belong to misidentifies every sample of the species it really is. Because Mash's winner-take-all strategy credits each shared hash to a single reference, such a reference only has to win by a few hashes to take them all, which also pushes the correct references far down the hit list: a runner-up at a much lower identity is not evidence that it matched poorly.

NCBI's own ANI check is where to start looking, but it does not settle the question on its own. The average_nucleotide_identity block of a NCBI Datasets assembly report carries match_status:

• species_match — the best-matching type assembly is the declared species.
• below_threshold_match — the best match is the declared species but below the species threshold, usually a divergent or undescribed lineage rather than a mislabel.
• below_threshold_mismatch — common and usually benign where species sit above 98% ANI or the declared species has no type assembly, as across the M. tuberculosis complex.
• mismatch — the best-matching type assembly is a different species. Worth investigating, though most hits are genus reclassifications (Mycobacterium → Mycolicibacterium) or reciprocal synonym pairs: of 3,534 references in mycobacteriaceae, 46 were flagged and 2 were real.

Confirm before relabelling, with evidence that suits the group. ANI against the type strain of both candidate species decides it for most genera; within the M. tuberculosis complex, where ANI is uninformative, the regions of difference do (GCA_013697125.1 was settled by the bovis-exclusive RD4 deletion). Record the correction, the deposited name and the evidence in curation/label_corrections.tsv, then apply it.

On an existing database, rewrite the sidecar and bins.tsv in place — no re-sketching, and --dry-run first:

scripts/apply_label_corrections.py curation/label_corrections.tsv \
    ~/.local/share/mashID/db/<db>.metadata.tsv <build_dir>/bins.tsv

Add --rank subspecies or --rank serovar when the bins.tsv was binned below species, so the corrected row's bin keeps the rank the others have. A <db>.confusable.tsv can be a target too.

Identification and TaxID are then correct. Description comes from the reference's own fasta header by way of the sketch, so it keeps the deposited name until the database is rebuilt.

On a rebuild, pass the same file to both steps. The binner needs it too: binning is what decides which genomes a reference is dereplicated against, and a mislabelled genome put in the wrong bin is compared only with the species it is named for, so it survives as a near-duplicate of the cluster it actually belongs to — which is how both mycobacteriaceae references got in.

scripts/bin_by_species.py assembly_data_report.jsonl ncbi_dataset/data binned/ \
    --corrections curation/label_corrections.tsv
make_mashID_db -i binned/ -o /path/to/db -p my_database \
    --assembly-report assembly_data_report.jsonl \
    --metadata curation/label_corrections.tsv

Choosing a sketch size

The pre-built taxon databases use s=10000, and the two 2025 ones (mycobacteriaceae-2025, listeria-2025) s=1000, which resolves species but yields coarse identities in steps of 0.001. For a custom database aimed at close relatives, s=10000 is worth the tenfold larger file: identities become finer and the shared-hash counts more informative.