Hugging Face Bio releases Carbon-A and 566 million predicted gene candidates

On 8 October 2026, Hugging Face Bio published Carbon-A: Finding genes in known and unknown genomes, by Georgia Channing, Leandro von Werra, and the Gener Team. The release is Carbon-Annotator (Carbon-A) plus a Carbon Annotation Database.

Sequencing reads the bases of a chromosome, the A, C, G, and T letters. Genome annotation is the next job: marking where genes sit, which stretches encode proteins, and how those coding stretches split into exons and introns. The post says sequencing now outruns that work, because annotation still depends on transcripts, known proteins, related species, and expert time, spread unevenly across the tree of life.

What the card says the model is

Hugging Face Bio describes Carbon-A as a 1.2-billion-parameter model that predicts protein-coding regions from DNA, one model for mammals, other vertebrates, invertebrates, plants, fungi, and protists. Context is 98,304 base pairs, on both strands. The model card matches that: 16,384 tokens of non-overlapping 6-mers, separate forward- and reverse-strand probabilities, default threshold 0.5. The licence field is mit. The card adds: “MIT. Apache-2.0 notices in the model code are retained.”

Training data on the card is RefSeq GCF assemblies for those eukaryotic groups (the card says protozoa; the post says protists). Bacteria and viruses are outside the training scope. Predictions are coding-sequence occupancy: transcript paths, isoforms, and full gene structures are not resolved. The card says: “The supplied evaluation panel was used during training; it is not a newly held-out test set.” Load it with trust_remote_code on Transformers 4.56 or later.

Predicted loci, in their numbers

The opening says the model was used “to discover 566 million new gene candidates” across 22,617 species. The database section is plainer: 48,167 assemblies, 22,617 taxa, about 27 trillion base pairs, and 566 million predicted protein-coding loci, about 11.0 times more taxa and 9.0 times more sequence than the RefSeq training corpus. Each row carries coordinates, a reconstructed coding sequence, a translated protein, and a confidence score. About half the target GenBank set is done, with another batch in three weeks. Call them predicted protein-coding regions, or gene candidates as the post does. They are not genes a lab has confirmed.

The benchmark is Hugging Face Bio’s too. On 42 genomes it reports a macro-averaged nucleotide F1 of 0.944, ahead of the baselines it lists: AUGUSTUS, Helixer, Tiberius, ANNEVO, OrionGeno, SegmentNT, and NTv3, at nucleotide, exon, and gene level. Individual baseline scores are not printed. A confidence score, it says, reaches an AUROC of 0.876 for exact coding-sequence matches. The post points to a technical report in the collection for a transfer test on plants.

Early Iso-Seq checks

Where Carbon-A and RefSeq disagree, Hugging Face Bio says it asked Active Site and UC San Diego to run PacBio Iso-Seq, which reads full RNA molecules from cells, on cat, Syrian hamster, chicken, and Arabidopsis. The lede names cats, chicken, and arabidopsis; the methods paragraph adds Syrian hamster. The post says Iso-Seq support is evidence a transcript structure is real, that several Carbon-A regions absent from RefSeq have that support, and that Carbon-A reaches roughly 0.62 complete-CDS support, close to the reference.

It then limits the claim. The experiments support transcription into RNA and the exon structure. They do not yet establish that those RNAs are translated into proteins, or what those proteins do. That is early validation of some predictions, in the post’s own next sentence. Ribosome profiling and mass spectrometry are named as later steps.

Fergal Martin, eukaryotic annotation team leader at EMBL-EBI, is quoted: “This new dataset was built on the huge number of publicly available genomes in the European Nucleotide Archive and could help researchers interpret and compare genes and genomes at scale.” The post says the annotations will also soon be available through EMBL.

The explorer searches by assembly, contig, or source key and shows strand-specific coding probabilities. Filter on the confidence score. The evaluation panel was already used in training.