sunlab-ai/ProRiboGen-dataset

ProRiboGen Dataset

Protein-RNA interaction and RBP motif dataset for ProRiboGen. The release covers 337 RNA-binding proteins (RBPs) and includes CLIP-derived RBP-RNA interaction pairs, a protein-level train/test split, reference motifs for held-out RBPs, model-generated RNA sequences, generated motifs, and motif-logo comparisons.

Quick access

Resource	Description
Training interactions	304 RBPs and 245,881 RBP-RNA interaction rows
Test interactions	33 held-out RBPs and 31,246 RBP-RNA interaction rows
Train/test split metadata	RBP lists, metadata, sequence identity, and structural-similarity checks
CLIP test FASTA files	CLIP binding-site sequences for the 33 test RBPs
CLIP motif matrices	CLIP-derived motif matrices for 26 test RBPs
Generated test FASTA files	Model-generated RNA sequences for the 33 test RBPs
Generated motif matrices	Motif matrices discovered from generated sequences
Motif-logo comparisons	Side-by-side CLIP and generated motif logos

Loading the dataset

The train and test interaction tables are configured as the default Hugging Face Dataset Viewer splits.

from datasets import load_dataset

dataset = load_dataset("sunlab-ai/ProRiboGen-dataset")

train_dataset = dataset["train"]
test_dataset = dataset["test"]

To download the complete repository, including FASTA files, motif matrices, logos, and split metadata:

hf download sunlab-ai/ProRiboGen-dataset \
  --repo-type dataset \
  --local-dir ProRiboGen_dataset

Repository structure

ProRiboGen_dataset/
├── 337_rbps_split/                 # Train/test RBP-RNA tables and split metadata
├── CLIP_test_fasta/                # CLIP binding-site sequences for test RBPs
├── CLIP_test_motif_matrix/         # CLIP-derived motif matrices for test RBPs
├── Generated_test_fasta/           # Model-generated RNA sequences for test RBPs
├── Generated_test_motif_matrix/    # Motif matrices from generated sequences
└── CLIP_Generated_test_logos/      # Side-by-side CLIP and generated motif logos

Interaction tables and split metadata

The 337_rbps_split directory contains the following files:

File	Description
ProRiboGen_train_RBP-RNA.csv	Training set containing 304 RBPs and 245,881 interaction rows
ProRiboGen_test_RBP-RNA.csv	Test set containing 33 RBPs and 31,246 interaction rows
train_test_RBPs.xls	Train/test split manifest, RBP metadata, and leakage checks

CSV columns

Column	Meaning
p_id	RBP identifier in Species-Gene format, for example Human-PUM1
protein	Amino-acid sequence
rna	RNA sequence containing A, C, G, and U
r_id	Transcript or binding-site identifier
s, t	Start and end coordinates of the binding interval
score	CLIP enrichment or binding score
type	Assay type, including PAR-CLIP, eCLIP, iCLIP, HITS-CLIP, and related assays
p_len, r_len	Protein-sequence and RNA-sequence lengths

Sheets in train_test_RBPs.xls

Sheet	Content
train_304 RBPs	Training RBP list with UniProt metadata and RNA counts
test_33 RBPS	Test RBP list with sequences and metadata
SeqId_Train vs Test	Pairwise sequence identity between training and test RBPs
TM-score_Train vs Test	Pairwise structural similarity between training and test RBPs

Motif and sequence resources

CLIP test FASTA files

The CLIP_test_fasta directory contains one FASTA file for each of the 33 test RBPs. These files contain CLIP binding-site RNA sequences used to construct reference motifs.

CLIP motif matrices

The CLIP_test_motif_matrix directory contains HOMER-style motif or position-weight matrix files with the .motif extension. Reference motifs were recovered for 26 of the 33 test RBPs.

The following RBPs have CLIP FASTA files but no CLIP motif matrix in this release:

• A.thaliana-HLP1
• Human-ALKBH5
• Mouse-Apc
• Mouse-Msi2
• Mouse-Pou5f1
• Mouse-Srsf2
• Mouse-Yy1

Generated test FASTA files

The Generated_test_fasta directory contains model-generated RNA sequences for the same 33 test RBPs, with one FASTA file per protein.

Generated motif matrices

The Generated_test_motif_matrix directory contains motif matrices discovered from the generated RNA sequences for all 33 test RBPs.

Motif-logo comparisons

The CLIP_Generated_test_logos directory contains PNG figures comparing CLIP-derived and generated motifs for the 26 RBPs with available CLIP motif matrices.

Train, validation, and test split

All interactions associated with a given RBP appear exclusively in either the training set or the test set. The two protein sets are disjoint, which prevents direct protein-level leakage.

Split	Proteins	Interaction rows	Fraction of proteins
Train	304	245,881	approximately 90%
Test	33	31,246	approximately 10%
Validation	None provided	None provided	Not applicable

Split constraints

The split was constructed using motif availability and anti-leakage criteria:

• Test RBPs with known or recoverable binding motifs were preferred.
• Every train-test protein pair has sequence identity of 40% or lower.
• Every train-test protein pair has a TM-score below 0.5.

Species composition

Split	Species composition
Test	Human 17, yeast 8, mouse 5, Arabidopsis thaliana 1, fly 1, worm 1
Train	Human 196, yeast 57, mouse 38, Arabidopsis thaliana 4, fly 4, worm 4, zebrafish 1

Assay composition

The interaction records are primarily derived from PAR-CLIP and eCLIP experiments. The dataset also contains iCLIP and HITS-CLIP records, together with smaller numbers of 4SU-iCLIP and Fr-iCLIP records in the training set.

Validation guidance

This release does not include an independent validation split. During model development, users should construct a protein-level validation split from the training RBPs. The provided test split should be reserved for final evaluation and should not be used for model selection or hyperparameter tuning.

Data sources and processing

1. RBP-RNA interaction tables
   The interaction records in 337_rbps_split/*.csv were curated from POSTAR3, which integrates CLIP-seq-derived RBP binding sites from PAR-CLIP, eCLIP, iCLIP, HITS-CLIP, and related assays across multiple species.

2. Protein metadata
   Protein identifiers, sequences, and associated metadata in the split workbook were aligned with UniProt records.

3. CLIP-derived motif resources
   The CLIP FASTA files and motif matrices were prepared from binding-site sequences associated with the held-out test RBPs.

4. Model-generated resources
   Generated FASTA files contain RNA sequences produced by the protein-conditioned ProRiboGen model. Generated motif matrices and comparison logos were produced through the project motif-analysis workflow.

Users should cite ProRiboGen, POSTAR3, UniProt, and the relevant primary CLIP data producers when reusing the corresponding components.

Limitations

• No independent validation split is included.
• Reference CLIP motifs were recovered for 26 of the 33 test RBPs.
• CLIP-derived records inherit the experimental biases, coverage differences, and assay-specific characteristics of their source studies.
• The repository contains components governed by different source terms, as described below.

License and terms of use

This repository contains materials from multiple sources and is therefore labeled with license: other.

The RBP-RNA interaction data are derived from POSTAR3 and remain subject to the usage terms stated by POSTAR3, together with the applicable terms of the corresponding primary data sources. UniProt-derived records remain subject to the applicable UniProt terms.

The ProRiboGen authors do not grant additional rights over third-party materials. Model-generated sequences, derived motif files, figures, and project-specific annotations are provided for scholarly research and reuse with appropriate attribution.

Users are responsible for reviewing and complying with the applicable source terms before redistributing or reusing third-party-derived components.

Citation

Please cite the formal ProRiboGen paper once it becomes available. Until then, the dataset may be cited as:

@misc{proribogen_dataset_2026,
  title        = {ProRiboGen Dataset},
  author       = {{Sun Lab}},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\\url{https://huggingface.co/datasets/sunlab-ai/ProRiboGen-dataset}},
  note         = {Protein-RNA interaction pairs and RBP motif resources for 337 RNA-binding proteins}
}

The source POSTAR3 database can be cited as:

@article{zhao2022postar3,
  title   = {POSTAR3: an updated platform for exploring post-transcriptional regulation coordinated by RNA-binding proteins},
  author  = {Zhao, Wenxian and Zhang, Shuyuan and Zhu, Yucheng and others},
  journal = {Nucleic Acids Research},
  volume  = {50},
  number  = {D1},
  pages   = {D287--D294},
  year    = {2022},
  doi     = {10.1093/nar/gkab702}
}