CodonMoE

CodonMoE is a Python package that implements Adaptive Mixture of Codon Reformative Experts (CodonMoE and CodonMoE-pro) for mRNA analyses.

Datasets

We include four public mRNA datasets, all bundled as CSVs in datasets/. Each file shares the same schema:

• Sequence: RNA sequence (A,U,C,G)
• Value: real-valued target
• Dataset: dataset identifier
• Split: train / valid / test

Dataset	File
MLOS	datasets/MLOS.csv
Tc-Riboswitches	datasets/Tc-Riboswitches.csv
mRFP Expression	datasets/mRFP_Expression.csv
CoV Vaccine Degradation	datasets/CoV_Vaccine_Degradation.csv

Installation

conda create -n codonmoe python=3.9 -y
conda activate codonmoe
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia -y 
git clone https://github.com/Kingsford-Group/codonmoe.git
cd codonmoe 
pip install -r requirements.txt 
pip install -e .

API Reference

CodonMoE

CodonMoE(input_dim, num_experts=4, dropout_rate=0.1)

Parameters:

• input_dim: Dimension of the input features
• num_experts: Number of expert networks in the Mixture of Experts
• dropout_rate: Dropout rate for regularization

CodonMoEPro

CodonMoEPro(
    input_dim,
    num_experts=4,
    kernel_num=100,
    kernel_sizes=(3, 4, 5),
    dropout_rate=0.1,
)

Parameters:

• input_dim: Dimension of the input features
• num_experts: Number of expert networks in the Mixture of Experts
• kernel_num: Number of convolutional kernels per kernel size
• kernel_sizes: Tuple of kernel sizes for multi-scale Codon Convolution
• dropout_rate: Dropout rate for regularization

CodonMoEPro is an enhanced version of CodonMoE.

mRNAModel

mRNAModel(base_model, codon_moe)

Parameters:

• base_model: The base model to be integrated with CodonMoE
• codon_moe: The CodonMixture of Experts model

API Tests

python -m unittest tests/test_codon_moe.py

Citation

If you find this repository useful, please consider citing our paper: CodonMoE: DNA Language Models for mRNA Analyses.

@article{du2025codonmoe,
  title={CodonMoE: DNA Language Models for mRNA Analyses},
  author={Du, Shiyi and Liang, Litian and Li, Jiayi and Kingsford, Carl},
  journal={arXiv preprint arXiv:2508.04739},
  year={2025}
}