Hugging Face Bio’s Carbon-A Marks 566 Million Gene Candidates Across 22,617 Species

Hugging Face Bio released Carbon-A and the Carbon Annotation Database on October 8, 2026, turning a 1.2-billion-parameter DNA model into a public map of protein-coding regions across much of GenBank’s eukaryotic tree. The blog post says the team used the model to discover 566 million new gene candidates across 22,617 species, including thousands that have never been studied in depth. The model weights ship under MIT on HuggingFaceBio/Carbon-A-1.2B; the database, explorer space and technical report sit in a HuggingFaceBio collection updated the same day.

Sequencing is no longer the bottleneck the post describes. Modern instruments produce assemblies far faster than labs can annotate them. Annotation still usually depends on transcript evidence, known proteins, interpolation from related species and expert review—resources that are unevenly distributed across the tree of life. Carbon-A is aimed at that gap: one shared model that predicts coding sequence directly from DNA for mammals, other vertebrates, invertebrates, plants, fungi and protists.

One model, nucleotide resolution, both strands

Carbon-A predicts protein-coding sequence at single-base resolution. It uses a non-overlapping 6-mer tokenizer, so each DNA token covers six bases, and a 16,384-token window that spans 98,304 base pairs. Two strand-specific classification heads emit separate forward- and reverse-strand probabilities, which matters when coding regions overlap on opposite strands. Binary calls use a configurable threshold that defaults to 0.5. Training data came from RefSeq GCF assemblies across those eukaryotic groups. The Transformers interface is AutoModelForTokenClassification with remote code trusted, for Transformers 4.56 or later.

On 42 benchmark genomes, Hugging Face Bio reports a macro-averaged nucleotide F1 of 0.944 and says Carbon-A outperformed the baselines it evaluated—AUGUSTUS, Helixer, Tiberius, ANNEVO, OrionGeno, SegmentNT and NTv3—at nucleotide, exon and gene level. An animal-only checkpoint still annotated plant genomes it had never seen in training. On Tetrahymena thermophila, which uses a nonstandard genetic code where two usual stop codons encode glutamine, the model held a 0.960 nucleotide F1 without a species-specific annotator. The pig genome offers a label-noise check: trained on a 2017 assembly, Carbon-A scored better against a newer 2026 assembly than against the older training labels, which the team reads as evidence it learned coding regularities rather than memorizing RefSeq quirks.

Wet-lab checks and a GenBank-scale database

Disagreements with reference annotations are the interesting cases. Working with Active Site and UCSD, the team ran PacBio Iso-Seq on cat, Syrian hamster, chicken and Arabidopsis samples. Iso-Seq reads full RNA molecules from living cells; where it supports a Carbon-A prediction missing from RefSeq, that is independent evidence the model may have found a real transcript structure rather than a false positive. In aggregate results cited in the blog, Carbon-A reaches roughly 0.62 complete-CDS support, matching the reference closely and beating other ab initio systems tested. Those experiments support transcription and exon structure; they do not yet prove translation or function. Ribosome profiling, mass spectrometry and targeted assays are listed as next steps. Additional wet-lab results are held for a later release.

The Carbon Annotation Database currently covers 48,167 assemblies from 22,617 taxa—about 27 trillion base pairs and 566 million predicted protein-coding loci. Relative to the RefSeq-derived training corpus, that is about 11.0× more taxa and 9.0× more sequence. Each prediction links to assembly, genomic coordinates, reconstructed coding sequence, translated protein and a confidence score. On the benchmark, that score reaches an AUROC of 0.876 for distinguishing exact CDS matches from other predictions, so users can trade coverage for precision. Hugging Face Bio says it has annotated roughly half of its target GenBank set and plans another batch in three weeks. EMBL’s European Bioinformatics Institute is quoted as expecting the annotations through EMBL channels as well.

Run limits and what is still missing

The model card is explicit about limits. Predictions are CDS occupancy, not full gene structures with alternative isoforms. Bacteria and viruses are out of scope. Ambiguous bases and incomplete final 6-mers are excluded from scoring. Full-length windows want fused SDPA kernels on recent NVIDIA GPUs; without them, attention materializes large matrices and needs more than 32 GB of extra GPU memory. The supplied evaluation panel was used during training and is not a newly held-out test set. Inference examples on the card were checked on an NVIDIA H100 with BF16, PyTorch 2.7.1 and Transformers 4.56.0, including a full yeast chromosome I record in overlapping windows.

Carbon-A arrives as both a model and a database, with MIT-licensed weights and documentation on Hugging Face. The release pairs a cross-eukaryote annotator with wet-lab checks on familiar species and a public CADB that multiplies annotated sequence far beyond the RefSeq slice used in training. Teams with genomes outside the database can run the checkpoint locally and annotate their own assemblies.