Learning the native-like codons with a 5'UTR and secondary RNA structure aided species-informed transformer model

Overview

TransCodon is a transformer-based model for cross-species codon optimization, integrating 5′UTR sequences, coding regions, species identifiers, and RNA secondary structure features. It enables zero-shot prediction of gene expression potential and supports regulatory sequence design.

📦 Features

🧬 Joint modeling of 5`UTR and CDS

🌍 Species-specific codon usage learning

🔬 RNA secondary structure information

🧪 Validated in heterologous expression scenarios

📁 Dataset Access All training, fine-tuning, and held-out evaluation datasets are available at:

🔗 Google Drive Dataset Folder

The pretrained model weights can be downloaded from:

🔗 huggingface

🛠Installation

1. Clone the repository

   git clone https://github.com/guyuehuo/transcodon.git cd transcodon

2. Set up environment

We recommend using conda or [virtualenv].

conda env create -f environment.yml
conda activate transcodon

🌍 Usage

1. Pretrain

Generate DNA sequences from amino acid sequences using a pretrained TransCodon model:

python pretraining.py \
    --train_data data/finetune/train.csv \
    --output_dir checkpoints/pretain_model \
    --epochs 5 \
    --batch_size 3 \
    --accumulate_gradients 6\
    --lr 2e-4 \
    --num_gpus 4\

2. Finetune

Finetune the pretrained model on a custom dataset (e.g., for codon optimization or other downstream tasks):

python fintune.py \
    --train_data data/finetune/fintune.csv \
    --output_dir checkpoints/finetuned_model \
    --pretrained_model checkpoints/transcodon.pt \
     --epochs 15 \
    --batch_size 3 \
    --accumulate_gradients 6\
    --lr 2e-4 \
    --num_gpus 2\

3. Infer

Given an input amino acid sequence and a specified host species, TransCodon generates a DNA sequence that conforms to the natural codon usage landscape of the target species. This enables codon optimization for heterologous expression while preserving biological realism.

python infer.py \
    --input_data ./test.csv \
    --output_file ./optimized_dna.csv \
    --model_checkpoint checkpoints/finetuned_model.pt

📊 Evaluation

We provide python scripts for evaluation on metrics like:

Codon Recovery Rate

Codon Similarity Index (CSI)

Codon Frequency Distribution (CFD)

GC content 

MFE energy

%MinMax and DTW score between natural and generated sequences

📄 Citation

If you use this work, please cite:

@misc{TransCodon2025,
  title={Learning the native-like codons with a 5'UTR and secondary RNA structure aided species-informed transformer model},
  author={Hu et al.},
  doi = {10.1101/2025.07.19.665668},
  journal = {bioRxiv},
  year = {2025},
  note = {Preprint}
}

📬 Contact

For questions or feedback, feel free to contact: 📧 gu-yuehuo@qq.com