matthiasblum/uniprotkb-interpro-ted

UniProtKB proteins, domain annotations and ProtT5 embeddings

Every protein in UniProtKB with its sequence, taxonomy, best AlphaFold DB model, protein family and domain annotations, and a per-protein ProtT5 embedding.

Built from InterPro 110.0 and UniProtKB 2026_03.

Dataset	Rows	Size	One row per
proteins	150,002,165	33 GB	protein
annotations	2,396,506,422	37 GB	annotation
embeddings	150,002,165	265 GB	protein

150,002,165 of the 150,006,383 proteins in the UniProtKB 2026_03 release are included.

Layout

Each dataset is split into 151 part-NNNNN.parquet files of 1,000,000 proteins (the last holds the remaining 2,165), covering the same accession ranges, so proteins/part-00001.parquet and embeddings/part-00001.parquet describe exactly the same proteins in the same order, and annotations/part-00001.parquet holds exactly their annotations. A part can therefore be used on its own, without downloading the rest.

Rows are sorted by protein_acc, so Parquet statistics let a reader skip whole files when filtering on an accession. manifest.json lists every part with its first and last accession and its row counts.

Join the datasets on protein_acc.

Schemas

proteins

Column	Type	Description
protein_acc	string	UniProtKB accession, e.g. A0A000. Unique, sorted
protein_id	string	UniProtKB entry name, e.g. A0A000_STRVD
protein_name	string	Recommended or submitted protein name
protein_db	string	Swiss-Prot (reviewed) or TrEMBL (unreviewed)
is_fragment	bool	Whether the sequence is a fragment
sequence_length	int32	Residues
sequence	string	Amino acid sequence
sequence_md5	string	Uppercase MD5 of sequence
taxon_id	int32	NCBI taxonomy identifier
organism	string	Scientific name of the organism
superkingdom	string	Bacteria, Archaea, Eukaryota or Viruses; nullable
proteome_id	string	UniProt proteome, e.g. UP000005640; nullable
afdb_model_id	string	Best AlphaFold DB model, e.g. AF-A0A000-F1; nullable
afdb_mean_plddt	float32	Mean pLDDT of that model, 0–100; nullable

superkingdom is the NCBI rank domain (formerly superkingdom). It is null for 5,153 proteins whose taxon has no domain in its lineage (e.g. uncultured and environmental organisms).

annotations

Column	Type	Description
protein_acc	string	UniProtKB accession
signature_acc	string	Accession of the signature, feature or domain that matched
signature_id	string	Its short name, where the source provides one; nullable
signature_name	string	Its description; nullable, and always null for InterPro-N
signature_type	string	domain, family, homologous superfamily, region, conserved site, repeat, coiled coil, disordered, ptm
signature_db	string	Source database, e.g. Pfam, CATH-Gene3D, TED
entry_acc	string	InterPro entry integrating the signature, e.g. IPR015424; nullable
entry_id	string	Its short name, e.g. PyrdxlP-dep_Trfase; nullable
entry_name	string	Its description; nullable
entry_type	string	Its type; nullable
source	string	How the annotation was produced: InterPro, InterPro-N, TED or TEDLH-lite
score	float64	Confidence of the match; meaning depends on source, see below. Null for InterPro
locations	list<struct<start:int32, end:int32>>	Matched ranges on the sequence

locations coordinates are 1-based and inclusive. A discontinuous domain appears as several segments in one row, so locations has more than one element.

Coordinates stay within 1..sequence_length for every source but TED, whose domains were defined on the sequence AlphaFold modelled, which can differ from the current UniProtKB entry. 8,615 TED rows (0.009%) therefore have an end beyond the protein's sequence_length — clamp or discard them where exact coordinates matter.

The entry_* columns are null for the 40.8% of annotations whose signature is not integrated into an InterPro entry, and for all TED and TEDLH-lite rows.

embeddings

Column	Type	Description
protein_acc	string	UniProtKB accession
embedding	fixed_size_list[1024]	Per-protein ProtT5 embedding, float16

from huggingface_hub import hf_hub_download
import numpy as np, pyarrow.parquet as pq

t = pq.read_table(hf_hub_download(
    "matthiasblum/uniprotkb-interpro-ted", "embeddings/part-00000.parquet", repo_type="dataset"))
vectors = np.stack(t["embedding"].to_numpy(zero_copy_only=False))  # (1000000, 1024) float16

Annotation sources

source	Rows	What it is
InterPro	1,250,689,077	Matches to InterPro member database signatures, plus sequence features
InterPro-N	912,933,524	Matches predicted by InterPro-N
TED	95,118,221	CATH domains from The Encyclopedia of Domains (TED)
TEDLH-lite	137,765,600	CATH domains assigned by HMM search against TEDLH-lite

Scores

score carries a different quantity for each source, and the values are not comparable across sources. There is no separate score-type column because source already determines it:

source	score	Range in this release	Better when
InterPro-N	model confidence	0 - 1	larger
TED	Qscore	0 - 100	larger
TEDLH-lite	E-value	0 - Z	smaller
InterPro	null	-	-

score is non-null on every InterPro-N, TED and TEDLH-lite row and null on every InterPro row. The InterPro member databases are deliberately left unscored.

For InterPro-N, which scores each matched region separately, score is the highest confidence among the match's locations. TED and TEDLH-lite provide one value per match.

Member database versions

The InterPro and InterPro-N rows can report signatures from different versions of the member databases. The InterPro matches come from InterPro 110.0 (September 2026). InterPro-N is a deep learning model trained on InterPro 105.0 data (April 2025), and so it reports InterPro 105.0 signatures: the member database versions behind an InterPro-N row are those of that earlier release.

Versions behind each source. A dash means the database is not covered by InterPro-N:

Database	InterPro 110.0	InterPro-N (105.0)
AntiFam	8.0	–
CATH-Gene3D	4.3.0	4.3.0
CDD	3.21	3.21
COILS	2.2.1	–
HAMAP	2026_01	2025_01
MobiDB-lite	4.0	–
NCBIFAM	20.0	17.0
PANTHER	19.0	19.0
Pfam	38.2	37.3
Phobius	1.01	–
PIRSF	3.10	3.10
PRINTS	42.0	42.0
PROSITE patterns	2026_01	2025_01
PROSITE profiles	2026_01	2025_01
SFLD	4	4
SignalP_Euk	6.0i	–
SignalP_Prok	6.0i	–
SMART	9.0	9.0
SUPERFAMILY	1.75	1.75
TMbed	1.0.2	–

A signature rebuilt between the two releases may match different regions, or a different set of proteins, under the same accession, so an InterPro-N row and an InterPro row for the same signature_acc are independent evidence rather than duplicates. InterPro-N rows also carry no description — signature_name is always null — so join on signature_acc to an InterPro row to label one.

TED

Domains were obtained from the AlphaFold Protein Structure Database API. Every domain with a CATH assignment is included and its Qscore is reported in score:

• 76,387,244 rows carry a full CATH homologous superfamily, e.g. 3.90.1150.10.
• 18,730,977 rows carry a CAT fold, e.g. 1.10.287.
• A further 33,033,498 domains in the source have no CATH assignment at all and are not included, as there is no accession to identify them by.

Because the domains come from AlphaFold structures, proteins without an AlphaFold DB model have no TED annotations.

TEDLH-lite

TEDLH is the TED Library of HMMs, a library of profile HMMs built from TED domains. Sequences were annotated against tedlh-lite, a reduced version of that library, with hmmsearch (-Z 150006383 -E 0.001) and the hits were filtered with cath-resolve-hits (--min-dc-hmm-coverage=80 --worst-permissible-bitscore=25). The resulting independent E-value is reported in score.

signature_acc combines the CATH superfamily with the representative domain, e.g. 3.40.640.10_2w8tA02; the superfamily is the part before the underscore.

PANTHER

PANTHER matches a family and, through an ancestral node, a subfamily. Where a subfamily was assigned, it is reported instead of the family, as the more specific annotation: signature_acc is then e.g. PTHR13693:SF102, and the family remains readable as the part before the colon. 103,418,478 of the 220,890,633 PANTHER rows (46.8%) are subfamily-level; the rest are family-level, either because no subfamily was assigned or because the match came from InterPro-N, which does not assign subfamilies.

Examples

Python

One subset at a time, streamed, with datasets:

from datasets import load_dataset

proteins = load_dataset("matthiasblum/uniprotkb-interpro-ted", "proteins", split="train", streaming=True)

Pulling a single part of each dataset avoids downloading the whole thing, and the parts are accession-aligned, so these two cover the same 1,000,000 proteins in the same order:

from huggingface_hub import hf_hub_download
import pyarrow.parquet as pq

proteins = pq.read_table(hf_hub_download(
    "matthiasblum/uniprotkb-interpro-ted", "proteins/part-00000.parquet", repo_type="dataset"),
    columns=["protein_acc", "organism", "sequence_length"])
annotations = pq.read_table(hf_hub_download(
    "matthiasblum/uniprotkb-interpro-ted", "annotations/part-00000.parquet", repo_type="dataset"))

DuckDB

Querying it in place with DuckDB (INSTALL httpfs; once):

-- proteins per superkingdom; reads neither sequences nor embeddings
SELECT superkingdom, count(*)
FROM 'hf://datasets/matthiasblum/uniprotkb-interpro-ted/proteins/*.parquet'
GROUP BY 1 ORDER BY 2 DESC;

-- annotations per source database
SELECT signature_db, count(*)
FROM 'hf://datasets/matthiasblum/uniprotkb-interpro-ted/annotations/*.parquet'
GROUP BY 1 ORDER BY 2 DESC;

-- high confidence TED domains only (Qscore is in score for source = 'TED')
SELECT signature_acc, count(*) AS domains
FROM 'hf://datasets/matthiasblum/uniprotkb-interpro-ted/annotations/*.parquet'
WHERE source = 'TED' AND score >= 75
GROUP BY 1 ORDER BY 2 DESC;

-- TED domains on well-modelled Swiss-Prot proteins
SELECT p.protein_acc, a.signature_acc, a.signature_name, a.score, a.locations
FROM 'hf://datasets/matthiasblum/uniprotkb-interpro-ted/proteins/*.parquet' p
JOIN 'hf://datasets/matthiasblum/uniprotkb-interpro-ted/annotations/*.parquet' a USING (protein_acc)
WHERE p.protein_db = 'Swiss-Prot'
  AND p.afdb_mean_plddt > 90
  AND a.source = 'TED';

-- confident InterPro-N predictions that InterPro itself did not match,
-- i.e. where the two sources disagree for the same signature
SELECT n.protein_acc, n.signature_db, n.signature_acc, n.score
FROM 'hf://datasets/matthiasblum/uniprotkb-interpro-ted/annotations/*.parquet' n
WHERE n.source = 'InterPro-N' AND n.score > 0.95
  AND NOT EXISTS (
    SELECT 1 FROM 'hf://datasets/matthiasblum/uniprotkb-interpro-ted/annotations/*.parquet' i
    WHERE i.protein_acc = n.protein_acc
      AND i.signature_acc = n.signature_acc
      AND i.source = 'InterPro');

-- naming an InterPro-N row, whose signature_name is always null
SELECT n.protein_acc, n.signature_acc, any_value(i.signature_name) AS name
FROM 'hf://datasets/matthiasblum/uniprotkb-interpro-ted/annotations/*.parquet' n
LEFT JOIN 'hf://datasets/matthiasblum/uniprotkb-interpro-ted/annotations/*.parquet' i
  ON i.signature_acc = n.signature_acc AND i.source = 'InterPro'
WHERE n.source = 'InterPro-N' AND n.protein_acc = 'A0A000'
GROUP BY 1, 2;

-- the domain architecture of one protein, in order along the sequence
SELECT signature_db, signature_acc, signature_name, score, locations
FROM 'hf://datasets/matthiasblum/uniprotkb-interpro-ted/annotations/*.parquet'
WHERE protein_acc = 'A0A000'
  AND signature_type = 'domain'
ORDER BY locations[1].start;

-- PANTHER subfamilies of a family, with the proteins assigned to them
SELECT signature_acc, count(*) AS proteins
FROM 'hf://datasets/matthiasblum/uniprotkb-interpro-ted/annotations/*.parquet'
WHERE signature_db = 'PANTHER'
  AND signature_acc LIKE 'PTHR13693%'
GROUP BY 1 ORDER BY 2 DESC;

-- how many Pfam domains sit on each human protein
SELECT p.protein_acc, count(*) AS pfam_domains
FROM 'hf://datasets/matthiasblum/uniprotkb-interpro-ted/proteins/*.parquet' p
JOIN 'hf://datasets/matthiasblum/uniprotkb-interpro-ted/annotations/*.parquet' a USING (protein_acc)
WHERE p.taxon_id = 9606 AND a.signature_db = 'Pfam'
GROUP BY 1 ORDER BY 2 DESC;

References

UniProtKB: The UniProt Consortium. UniProt: the Universal Protein Knowledgebase in 2025. Nucleic Acids Research (2025). doi:10.1093/nar/gkae1010

InterPro: Blum M, Andreeva A, Florentino LC, et al. InterPro: the protein sequence classification resource in 2025. Nucleic Acids Research (2025). doi:10.1093/nar/gkae1082

InterProScan 6: Blum M, Hobbs E, Florentino L, Bateman A. InterProScan 6: a modern large-scale protein function annotation pipeline. Bioinformatics Advances (2026). doi:10.1093/bioadv/vbag141

The Encyclopedia of Domains (TED): Lau AM, Bordin N, Kandathil SM, Sillitoe I, Waman VP, Wells J, Orengo CA, Jones DT. Exploring structural diversity across the protein universe with The Encyclopedia of Domains. Science (2024). doi:10.1126/science.adq4946

TEDLH: Alvarez Carreño C, Petrov AS, Waman VP, Sillitoe I, Orengo C. TEDLH: domain HMMs for sensitive detection of remote homologues. Bioinformatics (2026). doi:10.1093/bioinformatics/btag470