IDiom-DB

IDiom-DB-v1 is a dataset of 54M predicted intrinsically disordered protein regions (IDRs) curated from the AlphaFold Database (see Dataset curation) and used to train IDiom. This repository also contains generated, reference, and example sequences used in the preprint and IDiom GitHub repository.

Preprint   |   IDiom GitHub   |   IDiom Models

IDRs within full length proteins

The dataset contains 54,155,136 IDR records across three FASTA files in idiom-db/. Each record identifies an IDR and includes its available flanking protein context.

Split	IDR records	File	Size
Training	53,613,584	idiom-db-v1_train.fasta	25.3 GB
Validation	270,775	idiom-db-v1_validation.fasta	127.6 MB
Test	270,777	idiom-db-v1_test.fasta	127.7 MB

IDRs only

idiom-db/idrs_only/ contains the same records and split assignments as the context-containing FASTAs above, with each sequence cropped to its annotated IDR.

Split	IDR records	File	Size
Training	53,613,584	idiom-db-v1_train_idrs_only.fasta	6.8 GB
Validation	270,775	idiom-db-v1_validation_idrs_only.fasta	34.2 MB
Test	270,777	idiom-db-v1_test_idrs_only.fasta	34.2 MB

Sequence conventions

The IDR is marked in the FASTA header by _IDR_x-y at the end of the sequence identifier, before any whitespace. Here, x and y are the first and last residue positions of the IDR in the sequence below the header, counting from 1 and including both endpoints. Full-protein records also contain the flanking sequence outside these positions. For isolated IDR records, the marked region spans the entire sequence.

For example, this record marks QSSG as the IDR at positions 4–7:

>example_IDR_4-7
MEDQSSGACDE

For an isolated IDR, the header marks the entire sequence:

>example_IDR_1-4
QSSG

See the IDiom GitHub sequence conventions for loading these records with IDiom.

For IDR-only records, headers retain the original record identifier and append _IDR_1-L, where L is the extracted sequence length. For example:

>A0A1Y2LV99-F1_IDR_319-353_IDR_1-35
KKDEVKDGDKAEGANGEVKEGDKSKEKPKDEKKKP

The final _IDR_1-35 describes positions in the extracted sequence. The earlier _IDR_319-353 remains part of the accession and identifies the region in the source record. IDiom parses the last _IDR_ suffix, so these files work with its FASTA record reader. The retained source span refers to the source FASTA sequence, which may include only available protein context.

Dataset curation

IDiom-DB was curated from AlphaFold Database (AFDB) v4 as described in the preprint:

1. Cluster proteins. Cluster the 214M AFDB sequences with MMseqs2 at 90% sequence identity and 80% coverage, then identify IDRs in cluster representatives.
2. Identify disordered regions. Smooth AF2 pLDDT scores with a 15-residue averaging window. Scores below 70 indicate disorder, scores above 80 indicate folded regions, and intermediate scores define gaps. Folded or disordered runs shorter than 10 residues become gaps; gaps between disordered regions, or at a protein terminus next to a disordered region, are merged into the IDR. Other gaps are classified as folded.
3. Filter sequences. Discard IDRs shorter than 30 residues, proteins longer than 1,020 residues, and proteins that are entirely low-pLDDT. Remove records matching DisProt IDRs at ≥50% sequence identity and ≥80% bidirectional coverage.
4. Trim signal peptides. Use SignalP 6.0 to identify N-terminal signal peptides with cleavage-site probability ≥0.9. Trim these residues from overlapping IDRs and discard records whose remaining IDR is shorter than 30 residues.

The resulting dataset contains approximately 54M predicted IDRs with their available N- and C-terminal flanking contexts. A protein with multiple IDRs can contribute multiple records.

Download

Install the Hugging Face CLI:

pip install huggingface_hub

Download the full repository:

hf download jxliu2/idiom-db --repo-type dataset --local-dir idiom-db

Download a single split:

Training:

hf download jxliu2/idiom-db idiom-db/idiom-db-v1_train.fasta --repo-type dataset --local-dir idiom-db

Validation:

hf download jxliu2/idiom-db idiom-db/idiom-db-v1_validation.fasta --repo-type dataset --local-dir idiom-db

Test:

hf download jxliu2/idiom-db idiom-db/idiom-db-v1_test.fasta --repo-type dataset --local-dir idiom-db

Download all three IDR-only splits:

hf download jxliu2/idiom-db --repo-type dataset --include "idiom-db/idrs_only/*.fasta" --local-dir idiom-db

Download only the supporting sequence sets:

hf download jxliu2/idiom-db --repo-type dataset --include "other/*" --local-dir idiom-db

Downloads preserve the repository's directory structure under --local-dir.

Supporting sequences

Supporting sequence sets are in other/:

Directory	Contents	Size
other/generated_sequences/	IDiom generations, RL-SAE designs, baselines, controls, and parameter sweeps	306.4 MB
other/reference_sequences/	DisProt, CATH, compartment, effector/inert, and training-sample reference sets	8.1 MB
other/example_data/	Sequence inputs for the IDiom GitHub cookbook	1.0 MB

Versions

Version	Location
idiom-db-v1	This repository; used in the preprint
idiom-db-v0	Previous dataset repository

Sources

Please cite the preprint and the original sources for the data you use: AlphaFold DB, UniProt, DisProt, CATH, Kilgore et al. / ProtGPS, DelRosso et al., and Ginell et al..

License

IDiom-DB is distributed under CC BY 4.0. Retain original-source attribution and applicable terms when reusing source-derived data. IDiom code is MIT-licensed.