jheuschkel/SynCodonLM-V2-NoTokenType

Model Card for SynCodonLM

• This model is a replicate of that trained with species - token type ID, however, trained without any token type ID.
• This model is totally protein-agnostic, while the species-token type model may still have a small amount of spurious statistical focus.

---

Installation

git clone https://github.com/Boehringer-Ingelheim/SynCodonLM.git
cd SynCodonLM
pip install -r requirements.txt #maybe not neccesary depending on your env :)

Embedding a Coding DNA Sequence Using our Model Trained without Token Type ID

from SynCodonLM import CodonEmbeddings

model = CodonEmbeddings(model_name='jheuschkel/SynCodonLM-V2-NoTokenType') #this loads the model & tokenizer using our built-in functions

seq = 'ATGTCCACCGGGCGGTGA'

mean_pooled_embedding = model.get_mean_embedding(seq)
#returns --> tensor of shape [768]

raw_output = model.get_raw_embeddings(seq)
raw_embedding_final_layer = raw_output.hidden_states[-1] #treat this like a typical Hugging Face model dictionary based output!
#returns --> tensor of shape [batch size (1), sequence length, 768]

Citation

If you use this work, please cite:

@article{10.1093/nar/gkag166,
    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},
    journal = {Nucleic Acids Research},
    volume = {54},
    number = {5},
    pages = {gkag166},
    year = {2026},
    month = {02},
    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 six of seven benchmarks sensitive to DNA-level features, including messenger RNA 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.},
    issn = {1362-4962},
    doi = {10.1093/nar/gkag166},
    url = {https://doi.org/10.1093/nar/gkag166},
    eprint = {https://academic.oup.com/nar/article-pdf/54/5/gkag166/67103471/gkag166.pdf},
}
}