As always, we welcome feedback! Please get in touch and help us make RNAcentral even better! rnacentral.org/contact

Read more in our blog post: blog.rnacentral.org/2025/10/rnac... or our recent preprint: doi.org/10.1101/2025...

Gene identifiers will be stable across releases, even as new transcripts are added. Each gene gets metadata including RNA type and description from expert databases.
Find genes via text search, sequence pages, or download them in our GFF files.

We built 103,814 human ncRNA genes from 600,225 transcripts using machine learning + graph clustering. The pipeline was trained on Ensembl/GENCODE data and achieved 99.4% accuracy.
Most genes are lncRNAs (65,187), followed by antisense lncRNAs (16,790) and pre-miRNAs (8,560).

Why genes? Until now, transcripts differing by a single nucleotide were separate entries. For rRNAs, this meant thousands of nearly identical sequences with no established relationship.
Genes bring biological context and make it easier to find all variants of the same RNA.

13/12: Co-author update! @nanonancy.bsky.social was also instrumental in helping make sure the summaries were up to scratch! Thanks Nancy!

12/12 Big thanks to our co-authors @afg781.bsky.social, @antonipetrov.bsky.social, @alexbateman1.bsky.social and others! Read the full paper here: doi.org/10.1093/data... #bioinformatics #LLM #AI

11/12 But overall, this shows that with careful prompting and checking, LLMs can help address the curation bottleneck in bioinformatics! 🎯

10/12 Some limitations: We can only use open-access papers (highlighting the importance of #OpenAccess!), and LLMs sometimes struggle with complex information synthesis.

9/12 We've also made our entire dataset of contexts and summaries available:
huggingface.co/datasets/RNA...

8/12 Want to try it yourself? Search for RNAs with summaries at:
rnacentral.org/search?q=has...

7/12 All summaries are now available through @rnacentral.bsky.social - making it easier than ever to quickly understand what we know about specific RNAs

6/12 The results? We generated >4,600 summaries covering ~28,700 RNA transcripts! Expert evaluation showed 94% were rated good or excellent quality. 📈

5/12 The key innovation is our multi-stage checking system:

Reference validation
Automated fact-checking
Self-consistency verification
This helps ensure accuracy and proper attribution.

4/12 Our solution: Use GPT-4 with carefully designed prompts to read scientific papers and generate accurate summaries, complete with proper citations! 🤖

3/12 We focused on non-coding RNAs, where the curation gap is particularly acute. Most databases lack good summaries of what each RNA does, making it harder for researchers to quickly understand their function.

2/12 Why did we build this? Curation of scientific literature is becoming increasingly challenging. There's a growing gap between publication rates and the number of available curators.