mRNAutilus: Multi-Objective-Guided Discrete Generation of mRNA with Optimized Therapeutic Properties.

This repository describes mRNAutilus, an mRNA masked diffusion model for conditionally optimizing mRNA design in-silico.

System requirements

Hardware

This package requires a device connected with a CUDA-compatible GPU and driver for execution.

Software

This package is supported for Linux. The package has been tested on the following systems:

• Linux: Ubuntu 22.04

Python 3.11 and CUDA 12.4 are requirements -- see Installation below for python-specific package dependencies.

Installation

We use conda for environment management. env.yml contains all python-related dependencies needed to run this software.

# create conda environment
conda env create -f env.yml

# activate mrna environment
conda activate mrna

Data Availability

Pretraining and regressor data are available on Zenodo.

Pretraining

exp=lm/mdlm_150m
model_name=mdlm_150m
python train.py \
    experiment=${exp} name=${model_name} \

You can adjust the other training configurations in ~/configs/experiment/lm/mdlm_150m.yaml as needed.

Weights

Weights are available here: https://drive.google.com/file/d/1mU37sSkEPNClR0OYBpLLF66g44TJEL3i/view?usp=sharing

Sampling

The entrypoint is at generate_mcts.py, which can be run via: python3 generate_mcts.py. All arguments are controlled via the config.yaml file at root.

Embeddings

Embeddings can be retrieved similar to the following:

from src.mrna.lm.mdlm import Diffusion

model = Diffusion.load_from_pretrained(<ckpt>)
five_utrs = ['ACTG...', 'GTCA...']
cds = ['ATGGCAGGG', 'ATGGCAACA...']
three_utrs = ['ACTG...', 'GTCA...']
tokens = torch.tensor(model.net.alphabet.batch_tokenize(five_utrs, cds, three_utrs), dtype=torch.int64, device=device)

with torch.no_grad():
    with torch.autocast(device_type='cuda', dtype=torch.float16):
        hidden = model.net(tokens)['last_hidden_state']

License

This repository is under the CC BY-NC 4.0 license.

Citation

If you find this repository useful, please cite the following:

@misc{patel2026mrnautilusmultiobjectiveguideddiscretegeneration,
      title={mRNAutilus: Multi-Objective-Guided Discrete Generation of mRNA with Optimized Therapeutic Properties}, 
      author={Sawan Patel and Sophia Tang and Yesol Kim and Yinuo Zhang and Divya Srijay and Ping-Jung Lin and Shambhavi Shubham and Fengmei Pi and Cedric Wu and Sherwood Yao and Pranam Chatterjee},
      year={2026},
      eprint={2605.31296},
      archivePrefix={arXiv},
      primaryClass={q-bio.BM},
      url={https://arxiv.org/abs/2605.31296}, 
}