The Pulse
Hugging Face Opens Carbon-A Gene Finder for 22,617 Species
Hugging Face released Carbon-A, an open model for predicting protein-coding regions, alongside a database of 566 million candidate loci across 22,617 taxa. The company says early RNA sequencing supports some predictions missing from referen
AI.info Team · Written with AI ·
Hugging Face announced Carbon-A on October 8, pairing an open gene-prediction model with a database of 566 million candidate protein-coding loci. The database spans 22,617 taxa, including species with little or no prior study, the company said. The release targets a gap between the growing supply of sequenced genomes and the slower work of identifying which DNA regions encode proteins.
One model scans genomes across major groups
Carbon-A has 1.2 billion parameters and predicts protein-coding regions directly from DNA. Hugging Face says a single model covers mammals, other vertebrates, invertebrates, plants, fungi and protists. It processes sequence in windows of up to 98,304 base pairs and makes predictions at single-nucleotide resolution on both DNA strands.
The accompanying Carbon Annotation Database currently contains 48,167 genome assemblies, covering about 27 trillion base pairs. Each prediction comes with genomic coordinates, a reconstructed coding sequence, a translated protein and a confidence score. Hugging Face says the database has about 11 times as many taxa and nine times as much sequence as the RefSeq-derived training corpus; roughly half of the target GenBank assemblies have been annotated so far.
The company reports a macro-averaged nucleotide F1 score of 0.944 across 42 benchmark genomes and says Carbon-A beat seven evaluated gene-prediction systems at nucleotide, exon and gene level. That result needs qualification: the model card says the supplied evaluation panel was used during training, so it is not a newly held-out test set. The reported score therefore does not represent performance on an untouched benchmark.
RNA evidence supports some predictions
To check predictions that differ from RefSeq annotations, Hugging Face and collaborators at ActiveSite and the University of California, San Diego compared model outputs with PacBio Iso-Seq data from cat, Syrian hamster, chicken and Arabidopsis. The company says the RNA data support several predicted coding regions absent from RefSeq. It reports roughly 0.62 complete-CDS support in the aggregate comparison, close to the reference annotation and above the other ab initio systems it tested.
That evidence supports the presence and structure of transcripts; it does not show that the RNA is translated into protein or establish what any predicted protein does. Hugging Face identifies ribosome profiling, mass spectrometry and targeted functional experiments as next steps. Fergal Martin, Eukaryotic Annotation Team Leader at EMBL’s European Bioinformatics Institute, said: Such work highlights the different opportunities for applying AI to genome annotation.
Researchers can run the model themselves
Carbon-A is available under the MIT license and accepts GenBank or FASTA input. Researchers can download the model from Hugging Face and run the supplied inference scripts; the model card specifies a CUDA-enabled PyTorch environment. The weights download is 4.6 GB, and full-length inference relies on memory-efficient GPU attention kernels.
The model predicts whether individual DNA bases fall in coding regions, but it does not resolve full transcript paths or alternative isoforms. Its training and intended use are focused on eukaryotic sequences, not bacteria or viruses. Hugging Face plans another database release in three weeks and says it intends to continue validation, including experiments that could test whether candidate regions produce proteins.