Ab initio gene annotation model calling protein-coding sequence at base resolution across eukaryotes with one checkpoint, no species label.
No providers recorded yet. Browse all providers
Most genomes sequenced today will never get a curated annotation. Evidence-based pipelines need RNA-seq or protein homology that is scarce outside model organisms, and sequence-only ab initio predictors are usually trained per clade, so annotating a new fungus or protist means picking a checkpoint and hoping it transfers. Carbon-A (Carbon-Annotator), developed by Hugging Face and Beijing Zhongguancun Academy with collaborators at Tsinghua University, IBABI in Shenzhen, TIGEM and the University of Naples Federico II, replaces that choice with a single 1.2-billion-parameter encoder that reads raw genomic DNA and marks protein-coding sequence on both strands at single-base resolution, across mammals, other vertebrates, invertebrates, plants, fungi and protozoa. It takes no species label, clade-specific decoder or homology feature.
The authors call their training strategy "annotation as pre-training": rather than filtering the genome to curator-chosen regions for a self-supervised objective, Carbon-A is trained from random initialization to predict coding sequence over complete 98,304-bp windows, gene-free ones included, so any signal that helps locate coding sequence can be used. The model belongs to the same Carbon family as the decoder-only Carbon language models. It extends the dual-strand design of the prokaryotic GENERanno annotator to eukaryotes, using the 6-mer, long-context configuration of the GENERator backbone. The preprint was posted to bioRxiv on 21 October 2026 and has not been peer reviewed.
Carbon-A is a 26-layer bidirectional transformer encoder (hidden size 2,048, 32 query and 4 key-value heads, RoPE, SiLU) with a context of 16,384 tokens, or 98,304 bp. The released checkpoint holds 1.15 billion parameters. Training used 2,055 RefSeq assemblies from 2,047 taxa, 3.03 Tbp of sequence expanded by sliding-window augmentation to 44.27 million windows (4.35 Tbp). Smaller strides oversample fungi and protozoa, and random 0–5 bp offsets vary the 6-mer phase. Coding bases are under 2% of the corpus, so the positive class is weighted threefold in the loss. Training ran about 500,000 steps on 32 A100 GPUs. At inference, overlapping windows are averaged, a transition network and a grammar-constrained Viterbi decoder assemble gene structures, and the confidence network scores each gene.
On a 42-genome benchmark (28 reference genomes from the training snapshot and 14 assemblies added to RefSeq in 2026 and unseen by every model), Carbon-A achieved the highest overall nucleotide-, exon- and gene-level F1 among AUGUSTUS, Helixer, Tiberius, ANNEVO, OrionGeno, SegmentNT and NTv3, and ranked first on both subsets. The confidence score separates exact gene matches with an AUROC of 0.876; a 0.10 threshold gives gene-level precision 0.792, recall 0.691 and F1 0.736. Annotating human chromosome 1 takes 2,027 s on one GPU.
Carbon-A annotates eukaryotic assemblies that lack a curated gene set, turning a FASTA or GenBank file into GFF gene models with CDS and protein sequences; the confidence threshold sets the precision a project needs. The authors ran the model over all 48,167 eukaryotic GenBank assemblies (22,617 taxa, 27.18 Tbp) to build the Carbon Annotation Database (CADB), with 566.34 million predicted protein-coding loci, each linked to its taxon, coordinates, translated protein and confidence. That is about eleven times RefSeq's taxa, intended as training data for genomic and protein language models and as a gene-discovery resource.
Carbon-A shows that a single supervised encoder can give consistent, calibrated gene calls across eukaryotes without falling back on per-clade models. The authors tested its novel predictions with PacBio Iso-Seq transcriptomes from cat, Syrian hamster, chicken and Arabidopsis. Its overall Iso-Seq support rate (0.619) was close to RefSeq's (0.623), and 239 Carbon-A-specific loci absent from RefSeq had transcript support. Limitations are clear: the model predicts only protein-coding sequence, not UTRs, non-coding RNAs or regulatory elements, and it assigns no function. It is trained on the union of all annotated isoforms, while decoding emits one gene path per locus, so alternative transcripts are not resolved. The novel loci have transcript but not protein-level evidence, and CADB entries remain computational predictions. Weights and the annotation pipeline are released under Apache 2.0.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.