Summary
A bioRxiv preprint describes Carbon-A, a 1.2-billion-parameter model for predicting protein-coding genes across eukaryotes. Its authors used the model to create a database covering 48,167 genome assemblies and report strong performance across 42 benchmarks.
A bioRxiv preprint posted on October 21, 2026 describes Carbon-Annotator (Carbon-A), a 1.2-billion-parameter model that predicts protein-coding genes directly from genomic sequences. The authors say a single model checkpoint works across mammals, other vertebrates, invertebrates, plants, fungi and protozoa. They also used it to build a database of predicted annotations covering 48,167 eukaryotic genome assemblies.
Genome annotation identifies likely genes and their locations in a DNA sequence. For protein-coding genes, that includes identifying exons—the gene segments that contribute to the final RNA—and predicting the protein sequence they encode. Such predictions are useful for studying species whose genomes lack extensively curated annotations, but they are not all equivalent to experimentally confirmed genes.
One model for many branches of life
The paper describes conventional annotation as difficult to apply consistently across eukaryotes: supporting evidence is unevenly available among species, while many computational predictors are designed for particular taxonomic groups. Carbon-A is intended to address that gap by using one model across a broad range of eukaryotic genomes.
The authors evaluated it on 42 benchmark genomes. They report that Carbon-A achieved the highest overall F1 scores among the evaluated models at nucleotide, exon and gene levels. F1 combines precision and recall into a single measure; the three levels assess predictions at different scales, from DNA positions to exons and genes. The abstract also reports robustness in cross-domain tests and evaluations involving nonstandard genetic codes, as well as per-gene confidence scores that can help rank predictions by reliability.
A large catalogue of predicted genes
Applying Carbon-A to eukaryotic genome assemblies in GenBank produced the Carbon Annotation Database (CADB). The preprint reports that it currently contains 48,167 assemblies representing 22,617 taxa, with 27.18 Tbp of genomic sequence and 566.34 million predicted protein-coding loci, alongside their translated protein sequences. The authors compare its coverage with RefSeq, reporting approximately nine times as much sequence and eleven times as many taxa.
The scale could make a consistent set of gene predictions available for species that lack curated reference annotations. The database may also provide material for gene discovery and for training genomic and protein language models, as the authors propose. Its entries should be understood as model-generated predictions rather than a catalogue in which every locus has been individually confirmed.
Transcript evidence for candidate genes
To assess candidates not present in RefSeq, the researchers generated full-length PacBio Iso-Seq transcriptomes from four species. The preprint says these data provided independent experimental evidence for Carbon-A gene candidates absent from RefSeq. This offers a check beyond computational benchmark comparisons, while the reported transcript evidence concerns work in those four species—not experimental confirmation of every predicted locus in CADB.
The findings are presented in a preprint, and the abstract reports aggregate benchmark results rather than numerical F1 scores or a breakdown by species. Carbon-A and CADB therefore represent a broad annotation resource and a route to candidate discovery, with the strength of evidence varying between computational predictions and candidates supported by transcript data.