Hugging Face Bio open-sources Carbon-A, a 1.2B gene-finding model, and a database of 566 million predicted coding loci across 22,617 species; Iso-Seq checks in cat, chicken and Arabidopsis back some genes missing from RefSeq

Why it matters

【Key points】

On October 8, Hugging Face Bio released Carbon-A (Carbon-Annotator), a 1.2B-parameter DNA model that predicts protein-coding regions at single-base resolution on both strands over a 98,304-bp window, using one model across mammals, other vertebrates, invertebrates, plants, fungi and protists. Weights are Apache 2.0 on the Hub, with training data and a technical report. Across 42 benchmark genomes it reports macro nucleotide F1 of 0.944, ahead of AUGUSTUS, Helixer, Tiberius, SegmentNT, NTv3 and others at nucleotide, exon and gene level; an animal-only checkpoint still annotates unseen plants, and Tetrahymena (nonstandard genetic code) scores 0.960. Run over GenBank, the Carbon Annotation Database now holds 48,167 assemblies, 22,617 taxa, ~27 trillion bp and 566 million predicted coding loci, about half the target set, with EMBL to carry the annotations.

【Verifiable】

Blog, model card and explorer Space are public. With ActiveSite and UCSD, PacBio Iso-Seq on cat, Syrian hamster, chicken and Arabidopsis supports some Carbon-A coding regions absent from RefSeq (~0.62 complete-CDS support, close to RefSeq, above other ab initio tools). This shows transcription, not translation or function; 566 million are candidates, not confirmed genes.

【Our read】

1. The value is hypothesis generation at scale: annotation, not sequencing, is the bottleneck, and an open cross-species model with per-gene confidence (AUROC 0.876) plus a filterable database echoes the AlphaFold DB playbook.

2. Hugging Face explicitly cites closed models refusing legitimate biology questions and calls 2027 the year for bio: open players want to lock in base tools before frontier labs tighten bio access. Carbon-A only locates genes, not function or design, so its risk surface is small.

3. Watch for the promised wet-lab results on unstudied species and protein-level evidence (ribosome profiling, mass spec).