GENTEL-Lab/OpenRNA-v1-114M

--- 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--- 许可证：Apache-2.0 任务类别： - 文本生成 标签： - 生物学 - RNA - 基因组学 - 非编码RNA(non-coding RNA) - 编码RNA(coding RNA) - 转录组学 - FASTA(FASTA) 大小类别： - 100M<n<1B --- # OpenRNA-v1 OpenRNA-v1是一套经人工精选整理的数据集，包含**1.14亿条全长RNA序列**，涵盖生命之树所有主要RNA生物型。该数据集从17个公共数据库中收集并统一标准化，已被用于训练RNA语言模型(RNA language model)家族EVA。如需了解更多细节与使用示例，请参阅我们的论文或GitHub仓库。 ## 数据集统计信息 <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> - **总序列数**：114,186,538条（从约1.437亿条原始序列过滤得到） - **总核苷酸数**：2313亿条（是RNAcentral v25.0的7.7倍） - **覆盖范围**：所有主要RNA生物型——mRNA、长链非编码RNA(lncRNA)、微小RNA(miRNA)、核糖体RNA(rRNA)、转运RNA(tRNA)、核仁小RNA(snoRNA)、核小RNA(snRNA)、Piwi相互作用RNA(piRNA)、环状RNA(circRNA)、病毒RNA等 - **分类学覆盖范围**：细菌、古菌、真核生物、病毒 - **格式**：FASTA(FASTA) ## 分类学覆盖范围 OpenRNA-v1涵盖生命所有主要界，RNA序列采集自细菌、古菌、真核生物与病毒。下方的系统发育树展示了数据集中代表的生命之树各分支上的RNA生物型分布。 <p align="center"> <img src="imgs/Phylo_tree_rna_kingdom.svg" width="65%" style="background:white;" /> </p> ## 数据来源 OpenRNA-v1整合了来自17个公共数据库的序列： | 数据来源 | 类型 | 总序列数 | 过滤后序列数 | 占比 | 备注/组成 | |-------------|------|-----------|--------------|---|-------------------| | NCBI（NT与病毒库） | 综合型 | 63,007,114 | 56,232,788 | 49.25% | 包含RefSeq、GenBank与NCBI病毒数据集。 | | RNAcentral联盟 | 整合型 | 51,275,411 | 33,402,433 | 29.25% | 从RNAcentral、Rfam、GtRNAdb等聚合而来（不含SILVA）。 | | Ensembl数据库 | 基因组型 | 18,739,555 | 15,699,321 | 13.75% | 脊椎动物基因组（来自第114版的严格ENS编号）。 | | 环状RNA数据库 | 环状RNA(circRNA) | 2,219,291 | 2,132,252 | 1.87% | 从circBase、circAtlas等整合而来。 | | SILVA数据库 | 核糖体RNA(rRNA) | 1,312,521 | 552,442 | 0.48% | 高质量核糖体RNA数据集。 | | NONCODE数据库 | 长链非编码RNA(lncRNA) | 640,747 | 216,709 | 0.19% | 长链非编码RNA。 | | piRNAdb数据库 | Piwi相互作用RNA(piRNA) | 200,123 | 41,940 | 0.04% | Piwi相互作用RNA。 | | 其他来源 | 多样型 | 6,296,930 | 5,908,653 | 5.17% | 包含WormBase、FlyBase、snoDB、miRBase等。 | | **总计** | - | **143,691,692** | **114,186,538** | **100%** | 用于EVA模型训练的最终数据集。 | ## 数据整理流程 原始序列从上述来源收集后，将经过多阶段过滤流程： ### 核心与基因组档案 RNAcentral与Rfam作为结构骨架。对于Rfam，带有明确家族级证据的序列将被赋予特定功能标签（如snRNA、snoRNA），而缺乏高分辨率分类的短转录本（50–200 nt）将被归类为小RNA(sRNA)。整合Ensembl（341个物种）与NCBI NT以实现广泛的分类学覆盖；对于NCBI NT，使用`accession2taxid`进行精确的物种映射，并移除分类谱系不明的序列。 ### 专门功能RNA富集 长链非编码RNA的覆盖范围由NONCODE、LNCipedia、LncRNAWiki与LncRNADisease提供（已对LNCipedia条目去除poly(A)尾巴）。微小RNA从miRBase与MirGeneDB统一获取，明确区分成熟序列与前体序列。piRNAdb与snoDB分别覆盖Piwi相互作用RNA与核仁小RNA。环状RNA序列被标准化为线性表示以用于Token化(Token)。 ### 系统发育与病毒多样性 SILVA与PR2提供具有广泛分类学覆盖的核糖体RNA序列；SILVA的整理流程包括去除引物序列、载体污染与潜在嵌合体序列。NCBI病毒序列根据谱系元数据被分为6类病毒RNA类型（双链RNA(dsRNA)、正链单链RNA(ssRNA+)、负链单链RNA(ssRNA−)等）。 ### 去冗余 冗余处理分为两个阶段：(1) 数据库内精确去冗余：通过SHA256内容哈希实现；(2) 数据库间去冗余：使用`seqkit rmdup --by-seq`（xxHash算法）。通过严格的功能过滤保留16个核心RNA类别；歧义类别（如misc_RNA、假基因）与处理副产物（如微小RNA环区域）将被移除。 最终数据集包含**114,186,538**条唯一序列，各类RNA的组成如下： <p align="center"> <img src="imgs/data_pinchart.svg" width="60%" style="background:white;" /> </p> ## 数据处理流程 ### 序列聚类与采样 为在保留序列多样性的同时降低冗余，所有序列将通过**MMseqs2 easy-linclust**以50%的序列同一性（`MIN_SEQ_ID=0.5`）与80%的覆盖度（`COVERAGE=0.8`）进行聚类，最终得到**17,350,557个聚类簇**（其中56.19%为单序列簇，37.65%为包含2–10条序列的超小簇）。 训练样本采用**逆平方根加权采样**策略生成，其中第*i*个聚类簇的采样概率与1/√(聚类簇大小)成正比，以此抑制过度富集的家族，提升模型对稀有RNA类型的学习机会。 ### 序列标准化 1. 转换为大写形式 2. DNA到RNA的碱基转换（T→U） 3. 移除非标准碱基（仅保留A、U、G、C、N） 4. 添加5′/3′方向标记 5. 以50%的概率进行反向互补数据增强 ### Token化(Tokenization) 我们设计了双模式Token化方案，可支持生成式与填充式两种训练目标。 #### 因果语言模型(Causal Language Model, CLM)格式： <bos>|<lineage>;<rna_type>|5[SEQUENCE]3<eos> 示例：`<bos>|D__细菌;P__变形菌门;<rna_mRNA>|5AUGCUGC...3<eos>` #### 通用语言模型(Generalized Language Model, GLM)格式——支持区域重设计任务中的中间填充： <bos_glm>|<lineage>|AB<span_3>CD<span_17>EF<eos><span_3>XXX<span_17>YY<eos_span> 词汇表共包含**114个Token**，其中包括4种标准RNA碱基、方向标记、边界标记、15种RNA类型条件标记，以及GLM专用的片段标记（`<span_0>`至`<span_49>`）。训练过程中，分类学谱系前缀将被掩码（损失值=−100）。 ## 引用方式 若您在研究中使用OpenRNA-v1，请引用： 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} } 同时请酌情引用原始数据来源。主要参考文献如下： **RNAcentral**：RNAcentral联盟. RNAcentral in 2026: genes and literature integration. *Nucleic Acids Research*, 54(D1):D303–D313, 2026. **Rfam**：Kalvari I等. Rfam 14: expanded coverage of metagenomic, viral and microRNA families. *Nucleic Acids Research*, 49(D1):D192–D200, 2021. **MMseqs2**：Steinegger M与Söding J. MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. *Nature Biotechnology*, 35:1026–1028, 2017. ## 许可证 Apache 2.0