Protein Sequences · LiteFold
UniProtKB
Processed UniProtKB release — comprehensive, high-quality protein sequence and functional annotation knowledgebase (Swiss-Prot + TrEMBL) reformatted for ML pipelines.
What UniProtKB is
UniProtKB (the Universal Protein Knowledgebase) provides a comprehensive resource of protein sequences annotated with functional descriptions. This release, structured by LiteFold, reorganises the knowledgebase into uniform tabular and sequence files tailored for machine learning pipelines. It brings together manually curated records from Swiss-Prot, alternative splice isoforms, and computationally annotated records from TrEMBL.
The underlying database focuses on high-quality, non-redundant reference proteomes, combining literature curation with automatic annotation methods to predict information for unstudied proteins. By packaging these entries into deterministic train, validation, and test partitions across hundreds of shards, this processed version facilitates large-scale sequence modelling and bioinformatics research.
How it was made
The dataset aggregates 203,172,274 protein entries containing 75,747,523,712 amino-acid residues. These comprise 574,627 reviewed Swiss-Prot canonical sequences, 41,333 Swiss-Prot alternative isoforms, and 202,556,314 unreviewed TrEMBL sequences. The sequence data is stored across 205 zstandard-compressed FASTA shards, while parsed entry annotations are distributed across 615 gzip-compressed JSONL table shards alongside supplementary metadata records.
Protein-entry splits are determined using an exact-sequence hashing algorithm. This ensures that identical amino-acid sequences always fall into the same data split, producing 162,548,965 training records, 20,308,533 validation records, and 20,314,776 test records. A default index of 830 file entries is partitioned independently using a deterministic SHA-256 hash of each file identifier.
Key facts
- Publisher
- LiteFold
- Total protein records
- 203,172,274
- Total residues
- 75,747,523,712
- Swiss-Prot canonical records
- 574,627
- Swiss-Prot isoform records
- 41,333
- TrEMBL records
- 202,556,314
- Sequence shards
- 205
- Protein table shards
- 615
- Protein-entry train split
- 162,548,965 records
- Protein-entry validation split
- 20,308,533 records
- Protein-entry test split
- 20,314,776 records
- File formats
- JSONL, FASTA, JSON
What you can use it for
- Biological feature extraction
- Extract structural and functional feature representations from curated Swiss-Prot canonical sequences and alternative splice isoforms.
- Proteome-scale sequence streaming
- Stream and process subsets of TrEMBL or Swiss-Prot records across distributed compute environments without staging uncompressed archives locally.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from LiteFold/UniProtKB at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 733 rows of default/train
access_note
string_label- Default index over UniProtKB files. Load configs sprot, sprot_varsplic, or trembl for protein-entry rows.100%
compression
string_label- gzip74%
- zstd25%
- 1%
dataset_id
string_label- uniprotkb100%
download_pattern
string_label- sequences/sequence_uniprotkb_uniprot_trembl.fasta.gz/shard-*.fasta.zst25%
- tables/source_set=trembl/split=validation/*.jsonl.gz25%
- tables/source_set=trembl/split=train/*.jsonl.gz25%
file_id
string_text- min
- 9
- median
- 75
- mean
- 74.88
- max
- 87
is_metadata_records
bool- False100%
- True0%
is_sequence_shard
bool- False75%
- True25%
is_table_shard
bool- True74%
- False26%
path
string_text- min
- 9
- median
- 75
- mean
- 74.88
- max
- 87
records_in_source
int- min
- -1
- median
- 203M
- mean
- 199M
- max
- 203M
records_in_table_split
int- min
- -1
- median
- 20M
- mean
- 50M
- max
- 162M
records_total
int- min
- 203M
- median
- 203M
- mean
- 203M
- max
- 203M
repo_id
string_label- LiteFold/UniProtKB100%
residues_in_source
int- min
- -1
- median
- 75.5B
- mean
- 74.3B
- max
- 75.5B
residues_total
int- min
- 75.7B
- median
- 75.7B
- mean
- 75.7B
- max
- 75.7B
role
string_label- protein_entry_table_shard74%
- sequence_shard25%
- source_manifest0%
shard_index
int- min
- -1
- median
- 97
- mean
- 98.25
- max
- 202
shards_in_source
int- min
- -1
- median
- 203
- mean
- 199.68
- max
- 203
size_bytes
int- min
- 484
- median
- 11M
- mean
- 180M
- max
- 74.2B
source_file
string_label- sequence/uniprotkb/uniprot_trembl.fasta.gz98%
- sequence/uniprotkb/uniprot_sprot.fasta.gz1%
- 1%
source_set
string_label- trembl98%
- 1%
- sprot_varsplic1%
source_sha
string_label- ea4b6633410d5e2158cdb0b97fdcabd70bce3c27100%
source_slug
string_label- sequence_uniprotkb_uniprot_trembl.fasta.gz98%
- sequence_uniprotkb_uniprot_sprot_varsplic.fasta.gz1%
- 1%
split_bucket
int- min
- 1
- median
- 5
- mean
- 4.96
- max
- 9
table_split
string_label- 26%
- validation25%
- train25%
total_sequence_shards
int- min
- 205
- median
- 205
- mean
- 205
- max
- 205
Strengths and limits
Good at
- Deterministic exact-sequence hash partitioning prevents identical amino-acid sequences from leaking between training, validation, and test sets.
- Sharded across compressed JSONL and FASTA files to enable efficient indexing, parallel reading, and streaming.
Watch out for
- Protein-entry splits are based on exact sequences rather than sequence homology, necessitating external tools like UniRef or MMseqs for homology-aware benchmarking.
- The vast majority of entries come from TrEMBL and rely on automatic annotation rather than manual experimental verification.
Questions
Can I look inside UniProtKB without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is UniProtKB under?
Its authors declare cc-by-4.0. Read the terms on the original page before reusing the data.
What protein sources are included?
The dataset contains canonical reviewed proteins from Swiss-Prot (sprot), Swiss-Prot alternative splice isoforms (sprot_varsplic), and unreviewed automatic entries from TrEMBL (trembl).
How are the training, validation, and test splits divided?
The full protein tables use deterministic exact-sequence hash splits, guaranteeing that all occurrences of identical amino-acid sequences reside in the same split.
Are the data splits homology-separated?
No. The splits are strictly exact-sequence splits; strict homology-aware evaluations require creating custom splits using sequence clustering tools such as UniRef or MMseqs.
Can large subsets like TrEMBL be streamed rather than fully downloaded?
Yes, TrEMBL entries and specific table shards can be streamed sequentially or discovered through the default index table.
How to cite it
The UniProt Consortium (2025). UniProt: the Universal Protein Knowledgebase in 2025. Nucleic Acids Research, 53(D1), D609–D617. doi: 10.1093/nar/gkae1010
Written from UniProtKB's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.