Advancing Codon Language Modeling with Synonymous Codon Constrained Masking

• This repository contains code to utilize the model, and reproduce results of the paper Advancing Codon Language Modeling with Synonymous Codon Constrained Masking.
• Unlike other Codon Language Models, SynCodonLM was trained with logit-level control, masking logits for non-synonymous codons. This allowed the model to learn codon-specific patterns disentangled from protein-level semantics.
• Pre-training dataset of 43 Million CDS is available on Hugging Face here.

---

Installation

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

---

Usage

SynCodonLM uses token-type ID's to add species-specific codon context.

Before use, find the token type ID (species_token_type) for your species of interest here!

Or use our list of model organisms below

---

Embedding a Coding DNA Sequence

from SynCodonLM import CodonEmbeddings

model = CodonEmbeddings() #this loads the model & tokenizer using our built-in functions

seq = 'ATGTCCACCGGGCGGTGA'

mean_pooled_embedding = model.get_mean_embedding(seq, species_token_type=30) #E. coli
#returns --> tensor of shape [768]

raw_output = model.get_raw_embeddings(seq, species_token_type=30) #E. coli
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]

Codon Optimizing a Protein Sequence

This has not yet been rigourosly evaluated, although we can confidently say it will generate 'natural looking' coding-DNA sequences.

from SynCodonLM import CodonOptimizer

optimizer = CodonOptimizer() #this loads the model & tokenizer using our built-in functions

result = optimizer.optimize(
    protein_sequence="MSKGEELFTGVVPILVELDGDVNGHKFSVSGEGEGDATYGKLTLKFICTTGKLPVPWPTLVTTFSYGVQCFSRYPDHMKRHDFFKSAMPEGYVQERTIFFKDDGNYKTRAEVKFEGDTLVNRIELKGIDFKEDGNILGHKLEYNYNSHNVYIMADKQKNGIKVNFKIRHNIEDGSVQLADHYQQNTPIGDGPVLLPDNHYLSTQSALSKDPNEKRDHMVLLEFVTAAGITLGMDELYK", #GFP 
    species_token_type=30, #E. coli
    deterministic=True #true by default
)
codon_optimized_sequence = result.sequence

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},
}
}

---

Model Organisms Species Token Type IDs

Organism	Token-Type ID
E. coli	30
S. cerevisiae	118
C. elegans	212
D. melanogaster	190
D. rerio	428
M. musculus	368
A. thaliana	258
H. sapiens	373
C. griseus	345