ProteinLM-TDG2-Mutation

Code for paper: 📖Protein language models-assisted optimization of a uracil-N-glycosylase variant enables programmable T-to-G and T-to-C base editing

DOI: 🔗https://doi.org/10.1016/j.molcel.2024.01.021

@article{he2024protein,
  title={Protein language models-assisted optimization of a uracil-N-glycosylase variant enables programmable T-to-G and T-to-C base editing},
  author={He, Yan and Zhou, Xibin and Chang, Chong and Chen, Ge and Liu, Weikuan and Li, Geng and Fan, Xiaoqi and Sun, Mingsun and Miao, Chensi and Huang, Qianyue and others},
  journal={Molecular Cell},
  year={2024},
  publisher={Elsevier}
}

Install enviroment

Install the conda environment refering to the env.sh script.

Note that, DONNOT run the env.sh script directly. Instead, copy the commands in the env.sh script and run them in your terminal.

Download ESM series models

You can download the ESM series models from the huggingface model hub by searching facebook/model_name in the search bar.

Available models include:

Model name	link
esm1b_t33_650M_UR50S	link
esm1v_t33_650M_UR90S_1	link
esm1v_t33_650M_UR90S_2	link
esm1v_t33_650M_UR90S_3	link
esm1v_t33_650M_UR90S_4	link
esm1v_t33_650M_UR90S_5	link
esm2_t6_8M_UR50D	link
esm2_t12_35M_UR50D	link
esm2_t30_150M_UR50D	link
esm2_t33_650M_UR50D	link
esm2_t36_3B_UR50D	link
esm2_t48_15B_UR50D	link

Note that: For those who cannot access the huggingface model hub, you can download the models from the facebookresearch/esm.

We recommend you put the downloaded models in the Model directory, but this is not compulsory. No matter where you put the models, you should change the specified model path in the rank_all.py.

Create single mutations

First, rank your own data by feeding it into the rank_all.py script.

Note that you should change the --sequence-file and --output-file arguments to your own data.

python rank_all.py --sequence-file data/datadir/filename.fa --rank-strategy esm1v_5 --output-file data/datadir/filename_mutate_one_allsites.csv --model-name esm2_t48_15B_UR50D --single-site -f

Then, you can generate the landscape of your data by feeding it into the generate_landscape.py script.

python generate_landscape.py --mutate_one_sites data/datadir/filename_mutate_one_allsites.csv --figsize "(250,10)" --prefix filename --output data/datadir/filename_full_landscape.png

Benchmark in dms dataset

If you want to benchmark your model in the dms dataset, you can use the rank_all_dms.py script.

python rank_all_dms.py --dms-dir data/dms --model-name model_name --rank-strategy rank_strategy

For your convenience, we provide the rank_all_dms.sh script to run the benchmark in all models and rank strategies.

bash rank_all_dms.sh