Datasets:

cssbsnu
/

Thal-Kak_local_db

viewer: false
license: other
license_name: mixed-per-archive
license_link: >-
  https://huggingface.co/datasets/cssbsnu/Thal-Kak_local_db/blob/main/LICENSE.txt
language:
  - en
pretty_name: Thal-kak local MSA, template DB
size_categories:
  - 100B<n<1T
tags:
  - MSA
  - template
  - protein
  - NA

Thal-Kak local MSA, template databases

The sequence and template databases that the local MSA modes of Thal-Kak search — --msa mmseqs_local, --msa hhblits_local, --msa mmseqs_hhblits_local, and local template search on any of them.

Install these with install_db.sh, not by hand. Every file here is a multi-gigabyte .tar.zst holding a prebuilt MMseqs2 or HH-suite database; the installer verifies it, unpacks it into place and renames the files to the layout the pipeline expects.

At a glance


Archives	12 (`mmseqs/` 5, `hhblits/` 6, `template/` 1)
Download	664 GiB total
After extraction	2.31 TiB, plus 0.97 TiB of MMseqs2 search indexes built on your machine
Compression	zstd (`zstd` required to unpack)
Licenses	mixed, per archive — CC0 1.0 / CC BY 4.0 / CC BY-SA 4.0 / no grant (Logan). See License and attribution.

What is in this repository

These are derived databases, not upstream mirrors. Each archive was re-clustered and/or reformatted here, so the counts below are ours and will not match numbers computed on the upstream distribution. What was changed is recorded per source in LICENSE.txt, under Changes.

`mmseqs/` — MMseqs2 expandable profile databases

Searched by --msa mmseqs_local. Two layers: representatives are what the prefilter scans, members are what expandaln can pull into the alignment. The .idx search index is not shipped — install_db.sh builds it locally, which is where the third size column goes.

Archive	Download	Extracted	Index (local)	Representatives	Members
`uniref100_2026_01.tar.zst`	130.3 GiB	211.4 GiB	277.8 GiB	39,312,380	475,217,233
`mgnify_clusters.tar.zst`	126.9 GiB	202.3 GiB	470.8 GiB	313,067,917	717,738,164
`envhog_std.tar.zst`	20.7 GiB	36.7 GiB	53.6 GiB	11,742,979	129,896,064
`bfd_reduced.tar.zst`	18.6 GiB	27.6 GiB	114.7 GiB	51,652,611	65,984,053
`logan_human.tar.zst`	12.0 GiB	28.1 GiB	73.9 GiB	61,657,544	71,364,503

`hhblits/` — HH-suite (FFindex) databases

Searched by --msa hhblits_local. All but envhog were converted here from the corresponding MMseqs2 database, with cs219 and hhm profiles computed here; envhog is upstream's own HH-suite build.

Archive	Download	Extracted	Clusters (cs219)	Member sequences
`uniref100.tar.zst`	157.2 GiB	1,301.5 GiB	39,312,371	475,217,228
`uniref30_2023_02.tar.zst`	54.8 GiB	261.2 GiB	36,293,491	330,676,110
`mgnify_clusters.tar.zst`	33.7 GiB	150.3 GiB	31,779,642	278,185,288
`logan_nonhuman.tar.zst`	24.2 GiB	55.9 GiB	28,702,814	130,348,747
`envhog.tar.zst`	4.7 GiB	11.5 GiB	2,203,457	25,550,069
`logan_human.tar.zst`	0.8 GiB	2.0 GiB	1,645,263	7,385,713

`template/` — local template-search snapshot

Used when template.enable: true in the MSA config (the default on both local engines). A wwPDB/RCSB snapshot repackaged as a protein-only seqres FASTA, an MMseqs2 search database, per-entry gzipped mmCIF, a chain-metadata table and a deposition-date table.

Archive	Download	Extracted	Cutoff	PDB entries	Protein chains	Unique sequences
`BioMolDB_20260224.tar.zst`	80.5 GiB	81.2 GiB	2026-02-24	244,541	1,025,280	178,610

What is not here

install_db.sh can install four more databases, and fetches those from the upstream provider directly — nothing of them is redistributed here.

Installer command	Fetched from
`--family mmseqs uniref30_2302`	`opendata.mmseqs.org` (the ColabFold build)
`--family hhblits bfd`	`storage.googleapis.com/alphafold-databases` (full BFD)
`--family rna rfam`	EMBL-EBI
`--family rna rnacentral`	EMBL-EBI

Three consequences worth knowing before you plan a download:

Heteromers on hhblits_local also need the MMseqs2 uniref30_2302. The HH-suite UniRef100 carries no taxonomy, so multimer pairing is delegated to mmseqs, which needs that database's db_mapping / db_taxonomy sidecars.
logan_nonhuman exists in HH-suite format only. Its MMseqs2 build extracts to 1.4 TB and is not published anywhere. This asymmetry is deliberate.
RNA and RNP targets need the rna family, which is not here at all — it is built locally from Rfam and RNAcentral.

Getting the databases

With `install_db.sh` (recommended)

The installer resolves the archive, downloads it, verifies its SHA-256 against install/manifest/databases.tsv, extracts it into place atomically, normalizes the file stems to the layout db_paths.yaml expects, and builds the MMseqs2 .idx where one is needed. It is idempotent and resumable.

git clone https://github.com/CSSB-SNU/Thal-Kak.git
cd Thal-Kak
# create and activate the project environment first — it provides
# zstd, aria2c and mmseqs, which the installer calls.

./install_db.sh --family mmseqs                            # every mmseqs archive
./install_db.sh --family mmseqs uniref100_2026_01          # just this one
./install_db.sh --family hhblits envhog logan_human        # several by name
./install_db.sh --family template
./install_db.sh --family all                               # all four families
./install_db.sh --family mmseqs --status                   # what is already installed

To install somewhere other than <repo>/db, edit db_paths.yaml before running the installer — it reads that file and never writes it.

Manually

pip install -U "huggingface_hub[cli]"

hf download cssbsnu/Thal-Kak_local_db --repo-type dataset \
    --include "hhblits/envhog.tar.zst" "LICENSE.txt" --local-dir ./dl

--include filters the whole repo, so name LICENSE.txt alongside the archives — otherwise the terms that govern what you just downloaded stay behind on the Hub. A plain hf download <repo> --repo-type dataset takes everything and needs no such care.

Unpacking by hand is possible (zstd -dc <archive> | tar -x -C <dir>) but you then have to reproduce the layout yourself: each database lives at <family root>/<key>/, with every file inside renamed from its build-time stem to db (UniRef30_2023_02_a3m.ffdata → db_a3m.ffdata, and so on). install_db.sh does that rename for you. Prefer it unless you are mirroring.

Disk budget

Peak usage is higher than the final figure, because an archive sits on disk next to the tree being unpacked from it. Installing one family at a time keeps that overhead to a single archive.

Family	Archives here	Installed size
`hhblits`	6 of 7	1.74 TiB installed here (3.47 TiB with the full BFD)
`mmseqs`	5 of 6	1.46 TiB installed here (1.82 TiB with `uniref30_2302`)
`template`	1 of 1	81.2 GiB
`rna`	0 of 2	27.9 GiB, built locally

Everything, all four families: 5.40 TiB installed, ~5.7 TiB peak when installed family by family.

License and attribution

The Apache-2.0 license on the Thal-Kak source code does not apply to anything in this repository. Each archive carries the license of the source it derives from; no archive mixes sources. The authoritative record — upstream license, the license we place on our archive, attribution, citation, and exactly what we changed — is LICENSE.txt. The summary:

Archives	License we grant	Source
`mmseqs/uniref100_2026_01`, `hhblits/uniref100`	CC BY 4.0	UniProt UniRef100, release 2026_01
`hhblits/uniref30_2023_02`	CC BY-SA 4.0	UniRef30 2023_02, as built by ColabFold
`mmseqs/bfd_reduced`	CC BY-SA 4.0	BFD reduced subset, via the AlphaFold databases
`mmseqs/mgnify_clusters`, `hhblits/mgnify_clusters`	CC0 1.0	MGnify protein database, release 2024_04
`mmseqs/envhog_std`, `hhblits/envhog`	CC BY 4.0	EnVhogDB
`mmseqs/logan_human`, `hhblits/logan_human`, `hhblits/logan_nonhuman`	none	Logan v1 (SRA cutoff 2023-12-10)
`template/BioMolDB_20260224`	CC0 1.0	wwPDB / RCSB PDB, snapshot 2026-02-24

Three things that are easy to miss:

The Logan archives carry no license grant. No license instrument has been applied to that data by its distributor, so we grant no rights in it and make no representation about your right to use or redistribute it. It is passed through subject to the NCBI notice reproduced in LICENSE.txt, entry 6.
We do not license the output of the pipeline. An alignment or template file produced by a run contains material from whichever databases you searched. We make no claim over it either.
Two license claims are disputed upstream (uniref30_2302, bfd_reduced). Both disputes are written out in LICENSE.txt; we follow the more restrictive claim for anything we redistribute.

Everything here is provided AS IS, WITHOUT WARRANTY OF ANY KIND, either by the upstream provider or by us, and without any representation that it is fit for a particular purpose.

Citing

Cite the upstream database for whichever archive you used — each entry in LICENSE.txt carries its Cite line. For the pipeline itself, see the Thal-Kak repository.