---
license: apache-2.0
task_categories:
- text-generation
tags:
- biology
- RNA
- genomics
- non-coding RNA
- coding RNA
- transcriptomics
- fasta
size_categories:
- 100M<n<1B
---

# OpenRNA-v1

OpenRNA-v1 is a curated dataset of **114 million full-length RNA sequences** spanning all major RNA biotypes across the tree of life. Collected and harmonized from 17 public databases, OpenRNA-v1 was used to train the [EVA](https://huggingface.co/GENTEL-Lab/EVA) family of RNA language models. Please refer to our paper or GitHub repository for further details and usage examples.

## Dataset Statistics

<table align="center"><tr>
<td style="background:white;padding:8px 4px calc(5% + 8px) 8px;vertical-align:bottom;"><img src="imgs/data_comparsion.svg" width="100%" /></td>
<td style="background:white;padding:8px 8px calc(11.6% + 8px) 4px;vertical-align:bottom;"><img src="imgs/data_distribution.svg" width="100%" /></td>
</tr></table>

- **Total sequences**: 114,186,538 (filtered from ~143.7M raw sequences)
- **Total nucleotides**: 231.3 billion (7.7× larger than RNAcentral v25.0)
- **Coverage**: All major RNA biotypes — mRNA, lncRNA, miRNA, rRNA, tRNA, snoRNA, snRNA, piRNA, circRNA, viral RNA, and more
- **Taxonomic scope**: Bacteria, Archaea, Eukaryota, Viruses
- **Format**: FASTA

## Taxonomic Coverage

OpenRNA-v1 spans all major kingdoms of life, with RNA sequences sampled from Bacteria, Archaea, Eukaryota, and Viruses. The phylogenetic tree below illustrates the distribution of RNA biotypes across the tree of life represented in the dataset.

<p align="center">
<img src="imgs/Phylo_tree_rna_kingdom.svg" width="65%" style="background:white;" />
</p>

## Data Sources

OpenRNA-v1 integrates sequences from 17 public databases:

| Data Source | Type | Total Seqs | Filtered Seqs | % | Note / Composition |
|-------------|------|-----------|--------------|---|-------------------|
| NCBI (NT & Virus) | Comprehensive | 63,007,114 | 56,232,788 | 49.25% | Includes RefSeq, GenBank, and NCBI Virus datasets. |
| RNACentral Consortium | Integrated | 51,275,411 | 33,402,433 | 29.25% | Aggregated from RNACentral, Rfam, GtRNAdb, etc. (excl. SILVA). |
| Ensembl | Genomic | 18,739,555 | 15,699,321 | 13.75% | Vertebrate genomes (strict ENS IDs from Release 114). |
| CircRNA Databases | circRNA | 2,219,291 | 2,132,252 | 1.87% | Combined from circBase, circAtlas, and others. |
| SILVA | rRNA | 1,312,521 | 552,442 | 0.48% | High-quality ribosomal RNA datasets. |
| NONCODE | lncRNA | 640,747 | 216,709 | 0.19% | Long non-coding RNAs. |
| piRNAdb | piRNA | 200,123 | 41,940 | 0.04% | Piwi-interacting RNAs. |
| Others | Various | 6,296,930 | 5,908,653 | 5.17% | Includes WormBase, FlyBase, snoDB, miRBase, etc. |
| **Total** | - | **143,691,692** | **114,186,538** | **100%** | Final dataset for EVA training. |

## Data Curation

Raw sequences were collected from the sources above and subjected to a multi-step filtering pipeline:

**Core and Genomic Archives.** RNAcentral and Rfam serve as the structural backbone. For Rfam, sequences with definitive family-level evidence were assigned specific functional labels (e.g., snRNA, snoRNA), while short transcripts (50–200 nt) lacking high-resolution classification were curated as sRNA. Ensembl (341 species) and NCBI NT were integrated for broad taxonomic coverage; for NCBI NT, `accession2taxid` was used for precise species mapping and sequences with ambiguous taxonomic lineage were removed.

**Specialized Functional RNA Enrichment.** lncRNA coverage was provided by NONCODE, LNCipedia, LncRNAWiki, and LncRNADisease (poly(A) tails were trimmed from LNCipedia entries). miRNAs were unified from miRBase and MirGeneDB, with mature sequences explicitly distinguished from precursors. piRNAdb and snoDB covered piRNAs and snoRNAs respectively. circRNA sequences were standardized to linear representations for tokenization.

**Phylogenetic and Viral Diversity.** SILVA and PR2 provided rRNA sequences with broad taxonomic coverage; SILVA curation involved removing primer sequences, vector contamination, and potential chimeras. NCBI Virus sequences were classified into six viral RNA types (dsRNA, ssRNA+, ssRNA−, etc.) based on lineage metadata.

**Deduplication.** Redundancy was addressed in two stages: (1) intra-database exact deduplication via SHA256 content hashing; (2) inter-database deduplication using `seqkit rmdup --by-seq` (xxHash algorithm). A strict functional filter retained 16 core RNA categories; ambiguous categories (e.g., misc_RNA, pseudogene) and processing byproducts (e.g., miRNA loop regions) were excluded.

The final dataset comprises **114,186,538** unique sequences, with the following RNA type composition:

<p align="center">
<img src="imgs/data_pinchart.svg" width="60%" style="background:white;" />
</p>

## Data Processing

### Sequence Clustering and Sampling

To reduce redundancy while preserving diversity, all sequences were clustered using **MMseqs2 easy-linclust** at 50% sequence identity (`MIN_SEQ_ID=0.5`) and 80% coverage (`COVERAGE=0.8`), producing **17,350,557 clusters** (56.19% singletons, 37.65% ultra-small clusters of 2–10 sequences).

Training samples were drawn using an **inverse-sqrt weighted sampling** strategy where the sampling probability for cluster *i* is proportional to 1/√(cluster size), suppressing overrepresented families and improving model exposure to rare RNA types.

### Sequence Standardization

1. Uppercase conversion
2. DNA-to-RNA base conversion (T → U)
3. Removal of non-standard bases (only A, U, G, C, N retained)
4. Addition of 5′/3′ direction markers
5. 50% probability reverse complement augmentation

### Tokenization

We designed a dual-mode tokenization scheme supporting both generative and infilling training objectives.

**CLM (Causal Language Model) format:**
```
<bos>|<lineage>;<rna_type>|5[SEQUENCE]3<eos>
```
Example: `<bos>|D__Bacteria;P__Proteobacteria;<rna_mRNA>|5AUGCUGC...3<eos>`

**GLM (Generalized Language Model) format** — supports fill-in-the-middle for region redesign tasks:
```
<bos_glm>|<lineage>|AB<span_3>CD<span_17>EF<eos><span_3>XXX<span_17>YY<eos_span>
```

The vocabulary contains **114 tokens**, including 4 canonical RNA nucleotides, direction markers, boundary tokens, 15 RNA-type conditioning tokens, and GLM-specific span tokens (`<span_0>` to `<span_49>`). Taxonomic lineage prefixes are masked (loss = −100) during training.

## Citation

If you use OpenRNA-v1 in your research, please cite:

```bibtex
@article{huang2026eva,
  title = {EVA: A {{Generative Foundation Model}} for {{Universal RNA Modeling}} and {{Design}}},
  author = {Huang, Yanjie and Lyu, Guangye and others},
  year = {2026},
  journal = {bioRxiv},
  doi = {10.64898/2026.03.17.712398},
  url = {https://www.biorxiv.org/content/10.64898/2026.03.17.712398v1}
}
```

Please also cite the original data sources as appropriate. Key references:

**RNAcentral:** RNAcentral Consortium. RNAcentral in 2026: genes and literature integration. *Nucleic Acids Research*, 54(D1):D303–D313, 2026.

**Rfam:** Kalvari I, et al. Rfam 14: expanded coverage of metagenomic, viral and microRNA families. *Nucleic Acids Research*, 49(D1):D192–D200, 2021.

**MMseqs2:** Steinegger M & Söding J. MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. *Nature Biotechnology*, 35:1026–1028, 2017.

## License

Apache 2.0