tumorailab/protein2text-swissprot-split

Protein2Text Dataset.

This dataset contains the aggregated protein annotations for all SwissProt proteins as well as GSMC (microprotein database) and nmpfamsdb (metagenomic database). Swiss-Prot proteins are the only proteins that have a train/text split.

• 738,769 records across three sources
• 562,258 Swiss-Prot entries, 517,257 train / 45,001 test
• 4,796 homology-isolated proteins tagged as the dark-protein eval set (these are a subset of the Swiss-Prot test set)

Layout

records/
  swissprot/part-*.parquet     562,258 rows — carries `split` (train/test) + the eval flags
  nmpfamsdb/part-0000.parquet  100,000 rows
  gmsc/part-0000.parquet        76,511 rows
dark_set/
  isolated_4796_labels.tsv     the dark set with a per-label evidence breakdown
  label_inventory.md           what these 4,796 proteins are actually annotated with
figures/
  banded_seq30_max_similarity_violin.png
go_unmapped_names.tsv          GO names that could not be resolved to an id

Loading

from datasets import load_dataset

ds = load_dataset("tumorailab/protein2text-swissprot-split", "swissprot", split="full")

train = ds.filter(lambda r: r["split"] == "train")          # 517,257
test  = ds.filter(lambda r: r["split"] == "test")           #  45,001
dark  = ds.filter(lambda r: r["is_dark"])                    #   4,796

In pandas:

import pandas as pd
df = pd.read_parquet(
    "hf://datasets/tumorailab/protein2text-swissprot-split/records/swissprot/")
dark = df[df.is_dark.fillna(False)]

Split tags

Again, only present on Swiss-Prot records only; null on the metagenomic and microprotein sources, which are not part of any split.

column	values
split	train | test | null
is_dark, is_novel_family, is_novel_domain	true | false, and null when split != "test"

tag	n
split == "train"	517,257
split == "test"	45,001
is_dark	4,796
is_novel_family	1,308
is_novel_domain	490
none of the three	39,555

The three flags are independent and overlap: 836 dark proteins are also novel-family, 277 also novel-domain, 61 are all three.

How the split was built

I use a GraphPart style splitting method. A large similarity graph is constructed between all the Swiss-Prot proteins. I then interatively prune edges with stricter sequence similarity limits, resulting in sparser and sparser clsuters. At each stage, I assign certain clusters to train/test. This allows the test split to have a range of similarity to the train set (e.g. there are protein in the test set that are 90% similar to the train set and there are proteins that are 30% similar). This allows us to evaluate model performance while controlling on homology.

measured against the training set:

test group	n	median global identity	fraction >0.30 similarity
all test	45,001	0.49	68.3%
no flag set	39,555	0.55	77.6%
is_novel_domain	490	0.00	2.4%
is_novel_family	1,308	0.00	2.2%
is_dark	4,796	0.00	0.0%

The split — Swiss-Prot banded, seq30

How the dark label was assigned

In order to create a dark protein set that we can use to assess model performance on proteins (1) that have near-zero homology to the rest of the corpus and (2) still have relaible swiss-prot labels that we can eval on , we find singletons in the swiss-prot database that are dis-similar to the rest of the swiss-prot corpus.

1. All 562,258 Swiss-Prot sequences were searched all-against-all with mmseqs2. Here are the specific settings used: search -s 7.5 -e 1e-3 --min-seq-id 0.20 -c 0.0 --cov-mode 0 --alignment-mode 3.
2. A protein is dark if its mmseq2 returns no other homologous proteins for the specified protein.
3. There are 4,796 such proteins in the test split.

Record schema

field	type	notes
uniprot_id	string	primary key, unique across all 738,769 records
sequence	string	amino acid sequence
sequence_length	int64
split	string	train / test / null — Swiss-Prot only
is_dark	bool	homology-isolated. Test records only, null elsewhere
is_novel_family	bool	no Pfam family shared with train. Test records only, null elsewhere
is_novel_domain	bool	no InterPro domain shared with train. Test records only, null elsewhere
train_similarity	struct	max_global_identity, max_local_identity — max identity to the training set. Test records only, null elsewhere
functional_annotation	struct	uniprot_function_description (CC FUNCTION free text), catalytic_activity[], ec_numbers[], subcellular_location[], go_molecular_function[], go_biological_process[], go_cellular_component[]
sequence_features	struct	domains[], sites_and_motifs[], structural_features[], each with name, type, coordinates
interpro_entries	list<struct>	interpro_id, name, type, description
catalytic_mechanism	struct	M-CSA mcsa_enzyme_name, mcsa_description (enzymes only)
cross_refs	struct	pfam[], interpro[], pdb[], has_pdb
organism	struct	name, taxon_id, lineage[] of {taxid, name, rank}
source	string	swissprot | nmpfamsdb | gmsc

Source-specific extras, null elsewhere: division and domain (Swiss-Prot, 209,086 records); ecosystem[], family_id, has_structure, taxon[], provenance (NMPFamsDB and GMSC only).

Annotation coverage

	swissprot	nmpfamsdb	gmsc
records	562,258	100,000	76,511
median length	302	94	74
CC FUNCTION text	82.9%	0%	0%
GO molecular function	87.0%	0%	0%
GO biological process	81.9%	0%	0%
GO cellular component	82.2%	0%	0%
EC number	48.4%	0%	0%
subcellular location	62.8%	0%	0%
InterPro entries	97.3%	0%	0%