APCyc: Property-Informed Design of Cyclic Peptides via Automated Cyclization

Yifan Zhao, Lang Qin, Jintai Chen
The Hong Kong University of Science and Technology (Guangzhou)

Paper · HF paper page · Code · Project page

APCyc generates cyclic-peptide sequences and structures conditioned on a receptor pocket. It combines latent diffusion, cyclization prediction and optional guidance for affinity, proteolytic stability, permeability proxy, solubility and immunogenicity proxy.

Model files

Path	Use
weights/apcyc-legacy-vocab.pt	Generator bundle used in the example below
weights/apcyc-base.pt	Alternative generator bundle with the base input vocabulary
weights/property/vocab/	Five objective-specific property heads
weights/property/frozen-property-scorer.pt	Self-contained scorer for decoded-structure scoring and guidance
deployment/inference_reconstructed.yaml	Decoded-structure guidance configuration for the supplied scorer

Generator bundles contain the AE, model architecture, vocabulary and geometry calibration. A separate AE checkpoint is not needed for generation. For generator-feature guidance, the property heads must match the exact generator and feature implementation; see the guidance documentation.

Download and generate

Clone the code repository and follow the installation guide. Run the commands below from its root directory.

Sign in to Hugging Face and agree to the access conditions on the model page. Access is granted automatically. Authenticate the download client with the same account before downloading:

python -m pip install huggingface_hub
hf auth login
python - <<'PY'
from huggingface_hub import hf_hub_download

hf_hub_download(
    repo_id="qiaoqian0214/APCyc",
    filename="weights/apcyc-legacy-vocab.pt",
    local_dir="outputs/models",
    token=True,
)
PY
python scripts/make_example.py --out outputs/input
python -m apcyc.sample \
  --bundle outputs/models/weights/apcyc-legacy-vocab.pt \
  --input outputs/input/manifest.jsonl \
  --config configs/inference/default.yaml \
  --out-dir outputs/sample --device cpu --seed 2021
python scripts/inspect_predictions.py outputs/sample

The example creates synthetic input for checking the interface. For design experiments, provide real receptor pockets using the input manifest format. Use --device cuda for GPU inference. Generated structures require the relaxation and evaluation workflow before scientific interpretation.

Training and evaluation

See Training for training the AE, diffusion model and property heads, and Evaluation for property measurement and evaluation protocols.

License

The model weights are released under the MIT license. Source code and third-party components retain their original license notices; see Third-party components.

Citation

@misc{zhao2026apcyc,
  title={APCyc: Property-Informed Design of Cyclic Peptides via Automated Cyclization},
  author={Yifan Zhao and Lang Qin and Jintai Chen},
  year={2026},
  eprint={2606.12991},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2606.12991}
}