jheuschkel/clustered-cds-dataset

Dataset Card for Dataset Name

This is the dataset that was used to train SynCodonLM-V2. It is made up of more than 43 million CDS sequences (without introns), from over 35,000 species. The CDS were clustered at the 90% nt similarity level, per species, using MMSeqs2. The github of our model SynCodonLM can be found here:(https://github.com/Boehringer-Ingelheim/SynCodonLM) The Hugging Face of our learned model weights can be found here:(https://huggingface.co/jheuschkel/SynCodonLM-V2).

Citation

If you use this work, please cite:

@article {Heuschkel2025.08.19.671089,
    author = {Heuschkel, James and Kingsley, Laura and Pefaur, Noah and Nixon, Andrew and Cramer, Steven},
    title = {Advancing Codon Language Modeling with Synonymous Codon Constrained Masking},
    elocation-id = {2025.08.19.671089},
    year = {2025},
    doi = {10.1101/2025.08.19.671089},
    publisher = {Cold Spring Harbor Laboratory},
    abstract = {Codon language models offer a promising framework for modeling protein-coding DNA sequences, yet current approaches often conflate codon usage with amino acid semantics, limiting their ability to capture DNA-level biology. We introduce SynCodonLM, a codon language model that enforces a biologically grounded constraint: masked codons are only predicted from synonymous options, guided by the known protein sequence. This design disentangles codon-level from protein-level semantics, enabling the model to learn nucleotide-specific patterns. The constraint is implemented by masking non-synonymous codons from the prediction space prior to softmax. Unlike existing models, which cluster codons by amino acid identity, SynCodonLM clusters by nucleotide properties, revealing structure aligned with DNA-level biology. Furthermore, SynCodonLM outperforms existing models on 6 of 7 benchmarks sensitive to DNA-level features, including mRNA and protein expression. Our approach advances domain-specific representation learning and opens avenues for sequence design in synthetic biology, as well as deeper insights into diverse bioprocesses.Competing Interest StatementThe authors have declared no competing interest.},
    URL = {https://www.biorxiv.org/content/early/2025/08/24/2025.08.19.671089},
    eprint = {https://www.biorxiv.org/content/early/2025/08/24/2025.08.19.671089.full.pdf},
    journal = {bioRxiv}
}

More Detail

We curated a comprehensive dataset of CDS from the NCBI RefSeq database (25), encompassing nine major organismal groups: Archaea, Bacteria, Fungi, Invertebrate, Plants, Protozoa, Mammalian Vertebrates, Other Vertebrates and Viruses. To ensure broad taxonomic coverage and minimize sampling bias, we included only one representative CDS dataset per species, explicitly excluding sub-species and cell line-specific entries. This strategy helped prevent overrepresentation of well-studied organisms and ensured a more balanced view of codon usage across the tree of life. To prevent overlap of our train and test set, as well as reduce overfitting to highly repetitive sequences, we applied clustering per species at the nucleotide level using MMSeqs2 (26) with an identity threshold of 90% and a coverage threshold of 80%. This replicates the settings used to generate very popular biological pre-training datasets such as Uniref 90 (27), although we clustered at the DNA level rather than protein level to allow some extra synonymous variation even if proteins were more similar. This extra step seems to have been omitted from almost all other codon language models, besides CaLM (13), which clustered at the protein level at 40% identity using CD-HIT (28). We believe this is a critical step in removing redundancy in CDS datasets and preventing models from overfitting to highly paralogous sequences. Given the disproportionate abundance of CDS entries in certain groups (e.g., Bacteria and Other Vertebrate), we applied stratified sampling to balance the dataset, with an intentional emphasis on mammalian vertebrates to support downstream modeling objectives. Rather than stratifying against a total number of CDS within our dataset, we stratified to balance the total number of codons per group to be equal to, or no larger than our dataset’s mammalian sequences. We did this considering on average, mammalian CDS were much longer than some others such as bacteria. This emphasis reflects the relevance of mammalian systems in biotherapeutics and mRNA-based therapeutics, where codon optimization plays a critical role. In total, the final dataset comprised 43,526,014 CDS entries from 35,618 unique species. We randomly partitioned the dataset into 98% training (42,655,495 sequences) and 2% testing (870,519 sequences) subsets. This large-scale dataset provides unprecedented diversity for codon modeling and is approximately four times larger than that used in any previously published codon language model (12, 13, 15, 17). All sequences were validated to ensure they contained only canonical nucleotides (A, C, G, T), were divisible by three, contained no internal stop codons and had no 100% matches within the same species.