REAPS: An All-Atom Receptor-Aware Geometric Deep Learning Framework for De Novo Design of Linear and Macrocyclic Peptide Binders

REAPS is an all-atom receptor-aware geometric graph neural network for sequence design of linear and macrocyclic peptide binders.

It reframes peptide binder sequence design as a conditional sequence generation problem by treating the receptor as a fully observable all-atom context, enabling receptor-aware design of both linear and macrocyclic peptide binders.

🔧 Installation

We provide the following commands to reproduce the environment for running REAPS on Linux systems:

conda env create -f environment.yaml
conda activate REAPS
pip install -e .

⚡ Inference

We provide example inference commands for peptide binder sequence design from a given protein–peptide complex structure. For optimal performance, we recommend using a checkpoint trained with a relatively low coordinate noise level (such as 0.02 Å).

📦 Model checkpoints

Pretrained model checkpoints are available on Zenodo: https://zenodo.org/records/21978090.

Download the archive and extract it into the checkpoints/ directory:

mkdir -p checkpoints
tar -xvzf REAPS_ckpts.tar.gz -C checkpoints/

📏 Linear peptide binder sequence design

python inference.py \
  --pdb_file example/1T79.pdb \
  --peptide_chain_id B \
  --checkpoint_path checkpoints/REAPS_n0.02_linearFT.ckpt \
  --fasta_output_path outputs \
  --num_samples 8 \
  --temperature 0.2 \
  --mode linear

⭕ Macrocyclic peptide binder sequence design

python inference.py \
  --pdb_file example/2CK0.pdb \
  --peptide_chain_id P \
  --checkpoint_path checkpoints/REAPS_n0.02_cyclicFT.ckpt \
  --fasta_output_path outputs \
  --num_samples 8 \
  --temperature 0.2 \
  --mode cyclic

📌 Notes

• The input PDB file should contain both receptor and peptide chains, with at least two chains present.
• --peptide_chain_id specifies the peptide chain to be designed.
• Use a checkpoint that matches the selected design mode:
  • linear mode → checkpoint for linear peptide binder sequence design
  • cyclic mode → checkpoint for macrocyclic peptide binder sequence design

📕 Training

📦 Data preprocessing

The data preprocessing scripts used for training and testing are available in notebooks/.

The processed datasets are available on Zenodo: https://zenodo.org/records/19600852.

The downloaded archive is approximately 11 GB in size. After downloading, extract the archive with:

mkdir -p /path/to/your/data/root
tar -xvzf REAPS_datasets.tar.gz -C /path/to/your/data/root

🤖 Training scripts

REAPS uses a Hydra-based training pipeline built on PyTorch Lightning. Training is launched through the main training script together with configuration files under configs/.

🔥 Pretraining on multichain protein complexes

python train.py \
  --config-path configs \
  --config-name pre_training.yaml \
  paths.data_dir=/path/to/your/data/root \
  data=PPI_dataset \
  data.mode=pre-training \
  data.max_tokens_per_batch=14000 \
  model.is_fine_tuning=False \
  model.coords_noise_scale=0.02 \
  model.lr=1e-3 \
  trainer.max_epochs=200 \
  logger.wandb.offline=True

📏 Linear peptide binder sequence design fine-tuning

python train.py \
  --config-path configs \
  --config-name fine_tuning.yaml \
  pretrained_weights_path=/path/to/pre-training/ckpt \
  paths.data_dir=/path/to/your/data/root \
  data=PPI_dataset \
  data.mode=fine-tuning \
  data.max_tokens_per_batch=6000 \
  model.is_fine_tuning=True \
  model.coords_noise_scale=0.02 \
  model.lr=1e-5 \
  trainer.max_epochs=50 \
  logger.wandb.offline=True

⭕ Macrocyclic peptide binder sequence design fine-tuning

python train.py \
  --config-path configs \
  --config-name fine_tuning.yaml \
  pretrained_weights_path=/path/to/pre-training/ckpt \
  paths.data_dir=/path/to/your/data/root \
  data=CPCore_dataset \
  data.max_tokens_per_batch=12000 \
  model.is_fine_tuning=True \
  model.coords_noise_scale=0.02 \
  model.lr=5e-5 \
  trainer.max_epochs=50 \
  logger.wandb.offline=True

🧪 NK3R peptide binder design via iterative hallucination

We provide a Jupyter notebook that reproduces the NK3R peptide binder design pipeline described in the paper.

The notebook is located at: NK3R_hallu_pep_binder_design/NK3R_Xpep_binder_design_pipeline.ipynb

This notebook shows how REAPS is integrated into an iterative hallucination pipeline for peptide binder design, including scaffold hallucination, REAPS-guided sequence design, and iterative structural refinement.

Running this pipeline requires additional dependencies, including Boltz-2, PyRosetta, ColabFold, and HighFold.

🙏 Acknowledgements

The core implementation of this repository is built upon UniIF, with modifications and extensions for all-atom receptor-aware peptide binder sequence design.

The source code of UniIF is available at https://github.com/A4Bio/ProteinInvBench. We sincerely thank the authors for their valuable contributions to the community.

✏ Citation

If REAPS is relevant to your work, please consider citing our paper:

@article{qiu2026reaps,
  title   = {REAPS: An All-Atom Receptor-Aware Geometric Deep Learning Framework for De Novo Design of Linear and Macrocyclic Peptide Binders},
  author  = {Qiu, Yongkang and Qiao, Jingxin and Zhu, Lengjing and Du, Le and Yang, Shengyong and Zou, Jun},
  journal = {Journal of Chemical Information and Modeling},
  year    = {2026},
  volume  = {66},
  number  = {19},
  pages   = {12477--12491},
  doi     = {10.1021/acs.jcim.6c02057},
  url     = {https://doi.org/10.1021/acs.jcim.6c02057}
}