README.md

---
license: mit
models:
- jxliu2/idiom
---

# Official data repository for IDiom. 

## Associated paper link: https://doi.org/10.64898/2026.04.10.717777

## Associated GitHub repository: https://github.com/rotskoff-group/idiom

## Associated HuggingFace models: https://huggingface.co/jxliu2/idiom

## To download the entire dataset (187 GB), execute from IDiom root directory: 
```
hf download jxliu2/idiom-datasets \
  --repo-type=dataset --local-dir ./datasets
```

## If you only want the FASTA files containing the curated 37M IDRs (12 GB and 3 GB), run: 
```
hf download jxliu2/idiom-datasets \
  idr_datasets/training_sequences/AFDB_IDR_90_FIM_512_full.fasta \
  --repo-type=dataset \
  --local-dir ./datasets

hf download jxliu2/idiom-datasets \
  idr_datasets/training_sequences/AFDB_IDR_90_FIM_512_idrs.fasta \
  --repo-type=dataset \
  --local-dir ./datasets
  ```

## Description of files under the directory `idr_datasets/training_sequences`: 

- `AFDB_IDR_90_reps.fasta`: contains the 53M cluster representatives after the initial 214M full
length AFDB protein sequences are clustered at 90% identity, 80% coverage.
- `AFDB_IDR_90_alldata.h5`: contains 73M IDRs as extracted from the AlphaFold Database according to
the Tesei logic (see paper Methods), and after filtering for IDRs belonging to the 53M cluster
representatives identified in AFDB_IDR_90_reps.fasta. This HDF5 file contains the following
keys: <KeysViewHDF5 ['accession_ids', 'full_avg_plddt', 'full_length', 'full_-
seq', 'idr_end', 'idr_length', 'idr_plddt', 'idr_start', 'idrs']>.
- `AFDB_IDR_90_FIM_512.h5`: is created from AFDB_IDR_90_alldata.h5 by filtering out IDRs
whose full length sequences are longer than 512 residues. We also find that ∼ 1/3 of records in
AFDB_IDR_90_alldata.h5 are fully low-pLDDT sequences, and we filter out those sequences
because we find that they are not representative of intrinsically disordered proteins.
For the remaining 37M IDRs, we apply the fill-in-the-middle (FIM)
transformation as well as IDP data augmentation as mentioned in the Methods, and place
those records into AFDB_IDR_90_FIM_512.h5. We note that we represent the \<N\>, \<C\>, and
\<I\> tokens with 1, 2, and 3, respectively, in this HDF5 file as well as in the codebase. This
is the final file used for the precompute and pre-training steps.
- **`AFDB_IDR_90_FIM_512_full.fasta`: contains the 37M full length sequences (in correct order,
not FIM-transformed) contained in AFDB_IDR_90_FIM_512.h5. The fasta header contains
_IDR_X-Y where X and Y are the 1-indexed indices of the start and end (inclusive) of the
intrinsically disordered region.**
- **`AFDB_IDR_90_FIM_512_idrs.fasta`: contains only the sequences of the 37M intrinsically dis-
ordered regions in AFDB_IDR_90_FIM_512_full.fasta, without their surrounding context.**

## Description of files under the directory `idr_datasets/generated_sequences`:

- `Generated IDPs`: 100,000 unprompted intrinsically disordered proteins.
- `Generated IDRs`: 101,700 intrinsically disordered regions generated using 1,017 DisProt flank-
ing contexts prompts (100 generated IDRs per prompt).
- `Generated NPM1 IDRs`: 100,000 sequences generated using the NPM1 flanking context as
the prompt (UniProt: P06748).
- `Generated ProtGPS Sequences`: 10,000 IDPs generated from post-trained checkpoints. Post-
training was done to optimize ProtGPS localization scores for the four target compartments:
chromosome, nucleolus, P-body, and stress granule.