μProtein is a general framework designed to accelerate protein engineering by integrating μFormer, a deep learning model for accurate mutational effect prediction, with μSearch, a reinforcement learning algorithm tailored for efficient navigation of the protein fitness landscape.

For more details, please refer to our paper in Nature Machine Intelligence.

This repository contains the following components:

• pmlm/ – Protein language model pretraining
• mu-former/ – Fitness landscape modeling using the pretrained protein language model
• mu-search/ – Navigating the constructed fitness landscape oracle
• pretrained/ – Pretrained PMLM model checkpoint (stored using Git LFS).

For more details, refer to the respective README files:

• μFormer
• μSearch
• PMLM Pretraining

If you are using our code or model, please cite the following paper:

@article{sun2025accelerating,
  title={Accelerating protein engineering with fitness landscape modelling and reinforcement learning},
  author={Sun, Haoran and He, Liang and Deng, Pan and Liu, Guoqing and Zhao, Zhiyu and Jiang, Yuliang and Cao, Chuan and Ju, Fusong and Wu, Lijun and Liu, Haiguang and others},
  journal={Nature Machine Intelligence},
  pages={1--15},
  year={2025},
  publisher={Nature Publishing Group UK London}
}

This repository is licensed under the MIT License.

For questions or collaborations, please contact the authors via email or open an issue in this repository.