README.md

---
# For reference on dataset card metadata, see the spec: https://github.com/huggingface/hub-docs/blob/main/datasetcard.md?plain=1
# Doc / guide: https://huggingface.co/docs/hub/datasets-cards
license_name: creative-commons-attribution-4.0-international
repo: https://doi.org/10.5281/zenodo.15164650
tags:
  - cyclic-peptides
  - protein-design
  - AlphaFold2
  - hallucination
  - structural-biology
pretty_name: AfCycDesign Hallucinated Scaffolds
configs:
  - config_name: main
    data_files:
      - split: afcycpep_hallucinated
        path: data/afcycpep_data_hallucinated.csv
      - split: afcycpep_experimental
        path: data/afcycpep_data_experimental.csv
---

# Dataset Card for AfCycDesign

Hallucinated scaffolds used by AfCycDesign for cyclic peptide design.

## Dataset Details

Sets 7-16 of hallucinated peptide cif files and experimental CCDC structures.

### Dataset Description

This dataset contains hallucinated cyclic peptide scaffold structures (in CIF format) generated using AfCycDesign, a deep learning approach built on AlphaFold2 for de novo design of cyclic peptides. The scaffolds span peptide lengths of 7–16 residues and were generated through a hallucination procedure that simultaneously samples sequence and structure to produce well-ordered cyclic peptides. Structures in this dataset were filtered for high prediction confidence (pLDDT > 0.9) by AfCycDesign, representing 24,104 structurally diverse peptides predicted to fold into their designed conformations. These scaffolds serve as starting points for incorporating functional elements such as target-binding motifs for therapeutic applications.

- **Curated by:** Jie Chen / jiechen7 (at) uw.edu, Zachary Drake / zacharydrake (at) g.ucla.edu, Adriana Hernandez Gonzalez / ahgonzalez (at) ucdavis.edu, Akshaya Narayanasamy / akshayanarayanasamy (at) gmail.com, Lina Maria Pena Pardo / linamp (at) umich.edu
- **Language(s) (NLP):** Not applicable (structural biology data)
- **License:** MIT

### Dataset Sources

- **Repository:** https://zenodo.org/records/15164650
- **Paper [optional]:** Rettie, S.A., Campbell, K.V., Bera, A.K. et al. Cyclic peptide structure prediction and design using AlphaFold2. *Nature Communications* **16**, 4730 (2025). https://doi.org/10.1038/s41467-025-59940-7

## Uses

### Direct Use

These hallucinated scaffolds are intended for use as starting structures in cyclic peptide binder design pipelines. Researchers can graft known binding motifs (e.g., hot loops from protein–protein interfaces) onto these scaffolds and redesign the remaining sequence using tools such as ProteinMPNN and Rosetta to create cyclic peptide binders against therapeutic protein targets. The scaffolds can also be used for benchmarking cyclic peptide structure prediction methods or as seed structures for further computational design campaigns.

### Out-of-Scope Use

These scaffolds are computational design models and have not all been experimentally validated. They should not be used as experimentally determined structures. The dataset is not intended for direct therapeutic use without extensive experimental characterization, including structural validation, binding assays, and stability testing. The scaffolds are composed of canonical L-amino acids only and do not natively support non-canonical amino acid design, though post-hoc substitution is possible.

## Dataset Structure

Each entry in the dataset includes a PDB file for hallucinated data representing the 3D atomic coordinates of a hallucinated cyclic peptide scaffold. Scaffolds range from 7 to 16 residues in length and are organized by peptide size (sets 7–16). Each scaffold represents a unique structural cluster identified through torsion bin-based clustering, where each residue is assigned a bin (A, B, X, or Y) based on its φ, ψ, and ω backbone torsion angles. All scaffolds in this dataset were predicted to fold into their designed structures with high confidence (pLDDT > 0.9) by AfCycDesign.
Experimental entries include CIF files.

### Curation Rationale

The scaffolds were generated to provide a large, structurally diverse library of well-folded cyclic peptide backbones for downstream functional design. Prior physics-based methods (e.g., Rosetta kinematic closure) were computationally expensive and struggled with larger macrocycles (11–13 residues) without additional disulfide crosslinks. AfCycDesign's hallucination approach overcomes these limitations, enabling rapid enumeration of diverse cyclic peptide topologies including structures with short α-helices, β-sheets, and loop-only conformations.

### Source Data

https://zenodo.org/records/15164650

#### Data Collection and Processing

Scaffolds were generated using the AfCycDesign hallucination pipeline, which modifies AlphaFold2 with a custom cyclic offset matrix for relative positional encoding. For each peptide size (7–13 residues initially, later extended to 16), 48,000 hallucinated models were generated. Structures were clustered using a torsion bin-based approach, and the highest-confidence member from each cluster (pLDDT > 0.9) was selected. The hallucination is guided by losses that optimize prediction confidence metrics (pLDDT and predicted alignment error) and the number of intramolecular contacts. The pipeline was implemented within the ColabDesign v1.1.2 framework.

#### Who are the source data producers?

The data was computationally generated by researchers at the Institute for Protein Design, University of Washington, and collaborators at Harvard University. Key contributors include Stephen A. Rettie, Katelyn V. Campbell, Asim K. Bera, Sergey Ovchinnikov, and Gaurav Bhardwaj.

#### Who are the annotators?

Annotations are algorithmically generated by the AfCycDesign pipeline and AlphaFold2 confidence metrics.

#### Personal and Sensitive Information

This dataset does not contain any personal, sensitive, or private information. All data is computationally generated structural models of synthetic peptide sequences.

## Bias, Risks, and Limitations

The scaffolds are computational predictions and may not all fold as designed in experimental conditions. Validation were performed on a representative subset (8 X-ray crystal structures, all with RMSD < 1.0 Å to design models), but the full set has not been experimentally characterized. The hallucination approach is limited to the 20 canonical L-amino acids and does not natively incorporate D-amino acids or other non-canonical residues. For peptide sizes of 11+ residues, the 48,000 hallucination runs may not have fully sampled the available structural space, meaning additional unique scaffolds likely exist. The confidence metric (pLDDT) used for filtering, while strongly predictive of structural accuracy, is not a guarantee of experimental success.

### Recommendations

Users should be aware that most of these are computationally predicted structures, not experimentally determined ones.

## Citation

**BibTeX:**

```bibtex
@article{rettie2025cyclic,
  title={Cyclic peptide structure prediction and design using AlphaFold2},
  author={Rettie, Stephen A. and Campbell, Katelyn V. and Bera, Asim K. and Kang, Alex and Kozlov, Simon and Flores Bueso, Yensi and De La Cruz, Joshmyn and Ahlrichs, Maggie and Cheng, Suna and Gerben, Stacey R. and Lamb, Mila and Murray, Analisa and Adebomi, Victor and Zhou, Guangfeng and DiMaio, Frank and Ovchinnikov, Sergey and Bhardwaj, Gaurav},
  journal={Nature Communications},
  volume={16},
  pages={4730},
  year={2025},
  publisher={Nature Publishing Group},
  doi={10.1038/s41467-025-59940-7}
}
```

**APA:**

Rettie, S. A., Campbell, K. V., Bera, A. K., Kang, A., Kozlov, S., Flores Bueso, Y., De La Cruz, J., Ahlrichs, M., Cheng, S., Gerben, S. R., Lamb, M., Murray, A., Adebomi, V., Zhou, G., DiMaio, F., Ovchinnikov, S., & Bhardwaj, G. (2025). Cyclic peptide structure prediction and design using AlphaFold2. *Nature Communications*, *16*, 4730. https://doi.org/10.1038/s41467-025-59940-7

## Glossary

- **AfCycDesign**: The deep learning framework described in this paper that adapts AlphaFold2 for cyclic peptide structure prediction and design by introducing cyclic relative positional encoding.
- **pLDDT**: Predicted Local Distance Difference Test — a per-residue confidence metric from AlphaFold2, ranging from 0 to 1, where higher values indicate greater prediction confidence.
- **PAE**: Predicted Alignment Error — a metric from AlphaFold2 that estimates the error in the relative position of pairs of residues.
- **Hallucination**: A de novo design approach that simultaneously generates sequence and structure by optimizing AlphaFold2 confidence metrics starting from a random sequence.
- **Pnear**: A Rosetta-derived metric (0 to 1) indicating folding propensity; a value of 1 means the designed structure is the single lowest-energy conformation.
- **CIF**: Crystallographic Information File — a standard file format for representing 3D molecular structures.
- **Torsion bin clustering**: A method for grouping peptide structures by assigning bins (A, B, X, Y) based on backbone φ, ψ, and ω dihedral angles.

## More Information

The AfCycDesign code is implemented within the ColabDesign framework. For additional details on methods and supplementary data, see the full paper and its supplementary materials at https://doi.org/10.1038/s41467-025-59940-7.

## Dataset Card Authors

Jie Chen / jiechen7 (at) uw.edu, Zachary Drake / zacharydrake (at) g.ucla.edu, Adriana Hernandez Gonzalez / ahgonzalez (at) ucdavis.edu, Akshaya Narayanasamy / akshayanarayanasamy (at) gmail.com, Lina Maria Pena Pardo / linamp (at) umich.edu