The announcement
On October 8, 2026, Hugging Face's HuggingFaceBio team released Carbon-A (Carbon-Annotator), a 1.2-billion-parameter AI model that finds protein-coding genes directly in DNA sequence, along with the Carbon Annotation Database (CADB) built with it. The model is on Hugging Face under Apache 2.0.
The headline number: using Carbon-A, the team says it identified 566 million new gene candidates across 22,617 species, thousands of them never studied before.
Why this is hard
Sequencing a genome has become cheap, but annotating it (saying where the genes are) is still slow and laborious. Classic tools like AUGUSTUS often need per-species tuning and extra data such as RNA. That's why many public genomes in GenBank are sequenced but unannotated.
How Carbon-A works
- A single model for mammals, other vertebrates, invertebrates, plants, fungi and protists.
- It reads windows of 98,304 base pairs and makes nucleotide-level predictions on both DNA strands.
- Across 42 benchmark genomes, it reaches a macro-averaged nucleotide F1 of 0.944 and, per the team, beats AUGUSTUS, Helixer, Tiberius, ANNEVO, OrionGeno, SegmentNT and NTv3 at nucleotide, exon and gene level.
- It gives a confidence score for each predicted gene (AUROC of 0.876 for separating exact predictions from others), letting you trade coverage for precision.
The database
The current CADB covers 48,167 genome assemblies from 22,617 taxa, about 27 trillion base pairs and 566 million predicted coding loci. That's about 11 times more taxa than the RefSeq-derived training corpus. If your species isn't there, you can run the model straight from Hugging Face.
Caveats
- These are predictions, not confirmed genes. The team validated some candidates in the lab in well-studied species such as cat, chicken and arabidopsis, with partners Active Site and UCSD. Broader wet-lab results on unstudied genomes are still to be published.
- The benchmarks are the team's own.
- The model finds protein-coding regions; it doesn't say what each protein does.
Why it matters beyond biology
This is an example of open scientific AI with practical impact: agriculture (plant breeding), animal health, biotech and biodiversity can all benefit from genomes annotated at scale. For businesses, it also shows that small, specialized models can beat established tools on well-defined tasks.
For R&D teams
- Compare Carbon-A with your current annotation pipeline on a genome you already know well.
- Use the confidence score to prioritize which candidates to validate in the lab.
- Check CADB before annotating from scratch.
Sources
- Hugging Face blog (Oct 8, 2026): https://huggingface.co/blog/HuggingFaceBio/carbon-annotator-genbank-genome-annotation
- Hugging Face, HuggingFaceBio/Carbon-A-1.2B: https://huggingface.co/HuggingFaceBio/Carbon-A-1.2B
Want to apply specialized AI models to R&D at your company? KGA IT offers a free consultation: kga-it.com.