This repository present the protein design framework described in the paper. The fixed-backbone protein sequence design model (PSdesign) were built based on AlphaDesign. We also fine-tuned the ESM2 model to get a specialized version, ESM2_TadA. These codes are only available for non-commercial usage.

The following python packages are required.

• numpy, pandas
• pytorch
• torch_scatter
• ESM2
• biotite

$ python train_test_design_fp16_v13v3best2.py

Before runing this script, you need to prepare the dataset( using ESMFold to generated pdb files). Then set the path of checkpoint for the pre-tained model from step1. The esmfold_inference.py file is directly from ESM2.

$ python esmfold_inference.py -i seqdump_8e2p_top4791.fa -o esmfold/8e2p_top4791 > seqdump_8e2p_top4791.log
$ python train_test_design_fp16_v13v3best2_tadA.py

Prepare the dataset: Execute PREdata_TadA.py, providing the fasta file as input, and obtain pt files stored within the train_4791 and test_4791 directories.

$ python PREdata_TadA.py

In TadA_train.py, specify the paths for the train_4791 and test_4791 directories. Run the code for training. The training process will retain the optimal model based on the loss on the test set. Multi-GPU training is supported in the training process. Example usage:

$ CUDA_VISIBLE_DEVICES=0,1 python TadA_train.py

Before runing this script, you need to set the path of checkpoint for the fine-tuned model from step2.

$ python run_predict_proteindesign_v3_tadA.py -i 8e2p.chainA.pdb -o finetune_tada_8e2p

$ python generate_tada_sequences.py

Please cite the following article:

• Ye Yuan, Yang Chen, Rui Liu, Gula Que, Yina Yuan, Haihua Cai, Guipeng Li. An AI-designed adenine base editor. doi:https://doi.org/10.1101/2024.04.28.591233

Author: Dr. Guipeng Li

Email: guipeng.lee(AT)gmail.com