AMPainter

Painting Peptides with Antimicrobial Potency through Deep Reinforcement Learning

Overview

This framework consists of three modules:

• a policy network to assign the mutation sites
• a fine-tuned protein language model to replace the assigned residues
• a predictor named HyperAMP to evaluate the antimicrobial activity

Requirements

• pytorch
• DHG
• ankh
• transformers
• sklearn
• Levenshtein
• modlamp

Usage

1) HyperAMP solely

Train predictor

cd HyperAMP
python ankh_embedding.py
python tfidf.py
python train.py # or train_5folds.py

Predict

python ./HyperAMP/predict.py  # change input file path first

Length of a input peptide sequence should be less than 40.

The output of a input sequence is the transformed score based on its logMIC value to represent its antimicrobial activity.

2) Fine-tune Ankh

Data processing

(This step can be skipped for using ./data/maskseq1by1_0.2.csv directly)

python ./finetune/ankh_ft_data_1by1.py

Run

python ./finetune/ankh_ft_noval.py

3) AMPainter framework

Train agent network

python ./RLloop/reinforce.py

Evolve input sequences

Default params: lr = 1e-3, batch_size = 128, n_steps = 10, iterations = 8

python ./RLevolve/reinforce.py

4) Fitness function version

Evolve input sequences

python ./fitness/reinforce.py

Citation

@article{Dong2025AMPainter,
  author = {Dong, Ruihan and Cao, Qiushi and Song, Chen},
  title = {Painting Peptides With Antimicrobial Potency Through Deep Reinforcement Learning},
  journal = {Advanced Science},
  volume = {12},
  number = {43},
  pages = {e06332},
  keywords = {antimicrobial peptide, deep reinforcement learning, directed evolution, sequence design},
  doi = {https://doi.org/10.1002/advs.202506332},
  year = {2025}
}