PepTune: De Novo Generation of Therapeutic Peptides with Multi-Objective-Guided Discrete Diffusion 🧬🔮 (ICML 2025)

Sophia Tang*, Yinuo Zhang* and Pranam Chatterjee

This is the repository for PepTune: De Novo Generation of Therapeutic Peptides with Multi-Objective-Guided Discrete Diffusion 🧬🔮 published at ICML 2025. It is partially built on the MDLM repo (Sahoo et al. 2024).

PepTune leverages Monte-Carlo Tree Search (MCTS) to guide a generative masked discrete diffusion model which iteratively refines a set of Pareto non-dominated sequences optimized across a set of therapeutic properties, including binding affinity, cell membrane permeability, solubility, non-fouling, and non-hemolysis.

Environment Installation

conda env create -f src/environment.yml

conda activate peptune

Model Pretrained Weights Download

Follow the steps below to download the model weights required for this experiment.

1. Download the PepTune pre-trained MDLM checkpoint and place in checkpoints/: https://drive.google.com/file/d/1oXGDpKLNF0KX0ZdOcl1NZj5Czk2lSFUn/view?usp=sharing
2. Download the pre-trained binding affinity Transformer model and place in src/scoring/functions/classifiers/: https://drive.google.com/file/d/128shlEP_-rYAxPgZRCk_n0HBWVbOYSva/view?usp=sharing

Training Data Download

Download the peptide training dataset from https://drive.google.com/file/d/1yCDr641WVjCtECg3nbG0nsMNu8j7d7gp/view?usp=drive_link and unzip it into the data/ directory:

# Download peptide_data.zip into the data/ directory
cd data/

# Unzip the training data
unzip peptide_data.zip

cd ..

After unzipping, the data should be located at data/peptide_data/.

Repository Structure

PepTune/
├── src/
│   ├── train_peptune.py          # Main training script
│   ├── generate_mcts.py          # MCTS-guided peptide generation
│   ├── generate_unconditional.py # Unconditional generation
│   ├── diffusion.py              # Core masked discrete diffusion model
│   ├── pareto_mcts.py            # Pareto-front MCTS implementation
│   ├── roformer.py               # RoFormer backbone
│   ├── noise_schedule.py         # Noise scheduling (loglinear, logpoly)
│   ├── config.yaml               # Hydra configuration
│   ├── config.py                 # Argparse configuration
│   ├── environment.yml           # Conda environment
│   ├── scoring/                  # Therapeutic property scoring
│   │   ├── scoring_functions.py  # Unified scoring interface
│   │   └── functions/            # Individual property predictors
│   │       ├── binding.py
│   │       ├── hemolysis.py
│   │       ├── nonfouling.py
│   │       ├── permeability.py
│   │       ├── solubility.py
│   │       └── classifiers/      # Pre-trained scoring model weights
│   ├── tokenizer/                # SMILES SPE tokenizer
│   │   ├── my_tokenizers.py
│   │   ├── new_vocab.txt
│   │   └── new_splits.txt
│   └── utils/                    # Utilities & PeptideAnalyzer
│       ├── app.py
│       ├── generate_utils.py
│       └── utils.py
├── scripts/                      # Shell scripts for running experiments
│   ├── train.sh                  # Pre-training
│   ├── generate_mcts.sh          # MCTS-guided generation
│   └── generate_unconditional.sh # Unconditional generation
├── data/                         # Training data
│   ├── dataloading_for_dynamic_batching.py
│   └── dataset.py
├── checkpoints/                  # Model checkpoints
└── assets/                       # Figures

Pre-training

Before running, fill in HOME_LOC and ENV_LOC in scripts/train.sh and base_path in src/config.yaml to match your paths.

chmod +x scripts/train.sh

nohup ./scripts/train.sh > train.log 2>&1 &

Training uses Hydra configuration from src/config.yaml. Key settings:

• Backbone: RoFormer (768 hidden, 8 layers, 12 heads)
• Optimizer: AdamW (lr=3e-4, weight_decay=0.075)
• Data: 11M SMILES peptide dataset with dynamic batching by length
• Precision: fp64
• Checkpoints saved to checkpoints/ (monitors val/nll, saves top 10)

MCTS-Guided Peptide Generation

Generate therapeutic peptides optimized across multiple objectives using Monte-Carlo Tree Search.

1. Fill in base_path in src/config.yaml and src/scoring/scoring_functions.py.
2. Fill in HOME_LOC in scripts/generate_mcts.sh.
3. Create output directories: mkdir -p results logs

chmod +x scripts/generate_mcts.sh

# Usage: ./scripts/generate_mcts.sh [PROT_NAME] [PROT_NAME2] [MODE] [MODEL] [LENGTH] [EPOCH]
# Example: Generate peptides targeting GFAP with length 100
nohup ./scripts/generate_mcts.sh gfap "" 2 mcts 100 7 > generate.log 2>&1 &

Available Target Proteins

To specify a custom target protein, override +prot_seq=<amino acid sequence> and +prot_name=<name> as Hydra arguments in the generation script.

Scoring Objectives

PepTune jointly optimizes across five therapeutic properties via the integrated scoring suite:

Default MCTS Hyperparameters

These can be overridden via Hydra config overrides:

Unconditional Generation

Generate peptides without property guidance:

chmod +x scripts/generate_unconditional.sh

nohup ./scripts/generate_unconditional.sh > generate_unconditional.log 2>&1 &

Evaluation

To summarize metrics after generation, fill in path and prot_name in src/metrics.py and run:

python src/metrics.py

Citation

If you find this repository helpful for your publications, please consider citing our paper:

@article{tang2025peptune,
  title={Peptune: De novo generation of therapeutic peptides with multi-objective-guided discrete diffusion},
  author={Tang, Sophia and Zhang, Yinuo and Chatterjee, Pranam},
  journal={42nd International Conference on Machine Learning},
  year={2025}
}

License

To use this repository, you agree to abide by the Apache 2.0 License.