TransSAFP

About TransSAFP

TransSAFP is a self-attention based neural network that predicts antimicrobial label of self-assembling functional peptides.

Environments

The TransSAFP is based on TensorFlow 2.10 (with CudaToolkit 11.2). The bash commands below create a Conda environment for running TransSAFP inference.

conda create --name transSAFP
conda activate transSAFP
CONDA_OVERRIDE_CUDA="11.2" conda install tensorflow-gpu=2.10 cudatoolkit==11.2 -c conda-forge

Usage

To install, simply git clone this repository and activate the corresponding Conda environment.

To predict a native peptide sequence (without any unnatural amino acid or chemical modifications) using the pretraining model:

python run_pretrain.py {Your-Sequence-Here}
python run_pretrain.py AAAAAAAA # Predicts the antimicrobial activity label of the octa-alanine.

To predict a chemical modified peptide sequence using the transSAFP model:

python run_transSAFP.py {Your-Sequence-Here} {Your-Modification-Here}
python run_transSAFP.py AAAAAAAA C-HEX # Predicts the antimicrobial activity label of the octa-alanine with C-HEX N-terminal modification.

Available N-terminal types are:

- C8   
- C12  
- C16  
- PHE  
- BIP  
- DIP  
- NAP  
- ANT  
- PYR  
- C-PRO
- C-HEX

Description for files and directories.

• ./a_raw_datasets: Contains the N-terminal modified SAFPs data for the model training;
• ./b_preprocessing: Contains the tokenizers;
• ./c_transsafp: Contains the model weights;
• ./d_src: Contains the source codes for the TransSAFP model:
  • ./d_src/a_raw_datasets: Contains the N-terminal modified SAFPs data for the model training;
  • ./d_src/b_preprocessing: Contains the tokenizers and the transfer learning model and the preprocessed N-terminal modified SAFPs data;
  • ./d_src/c_transfer_learning: Contains the layers implemented for the TransSAFP model and the model weights;
  • ./d_src/map_seq8.py: The source codes for mapping the N-terminal octa-peptide space;
  • ./d_src/pretrain_model.py: The source codes for the pretraining model;
  • ./d_src/transfer_model.py: The source codes for the TransSAFP model;
  • ./d_src/transfer_model_eval.py: The source codes for evaluating the TransSAFP model;
  • ./d_src/*.ipynb: Analysis of the TransSAFP mode latent space;
• ./prediction.log: The prediction scores of the 140 SAFPs predicted and experimented synthesized in this study;
• ./run_pretrain.py: An interface for single sample inference of the pretraining model;
• ./run_transsafp.py: An interface for single sample inference of the TransSAFP model.

References

Huayang Liu†, Zilin Song†, Yu Zhang, Bihan Wu, Dinghao Chen, Ziao Zhou, Hongyue Zhang, Sangshuang Li, Xinping Feng, Jing Huang*, Huaimin Wang*,
De novo design of self-assembling peptides with antimicrobial activity guided by deep-learning,
Nature Materials, in press (2025).

Correspondence:
Prof. Huaimin Wang (wanghuaimin@westlake.edu.cn)
Prof. Jing Huang (huangjing@westlake.edu.cn)