RosettaCommons/AfCycDesign

--- # 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 ## Quickstart Usage ### Install HuggingFace Datasets package Each subset can be loaded into python using the Huggingface [datasets](https://huggingface.co/docs/datasets/index) library. First, from the command line install the `datasets` library $ pip install datasets then, from within python load the datasets library >>> import datasets ### Load dataset Load the 'RosettaCommons/AfCycDesign' datasets. >>> AfCycDesign = datasets.load_dataset('RosettaCommons/AfCycDesign') Downloading readme: 9.67kB [00:00, 3.57MB/s] Downloading data: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████| 2.94M/2.94M [00:00<00:00, 6.99MB/s] Downloading data: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████| 544/544 [00:00<00:00, 4.22kB/s] Generating afcycpep_hallucinated split: 100%|███████████████████████████████████████████████████████████████| 20656/20656 [00:00<00:00, 246500.54 examples/s] Generating afcycpep_experimental split: 100%|█████████████████████████████████████████████████████████████████████████| 8/8 [00:00<00:00, 3582.58 examples/s] and the dataset is loaded as a datasets.arrow_dataset.Dataset >>> AfCycDesign DatasetDict({ afcycpep_hallucinated: Dataset({ features: ['ID', 'sequence', 'nmer', 'rosetta_score', 'hf_path', 'Type'], num_rows: 20656 }) afcycpep_experimental: Dataset({ features: ['ID', 'sequence', 'nmer', 'rosetta_score', 'hf_path', 'Type'], num_rows: 8 }) }) which is a column oriented format that can be accessed directly, converted in to a pandas.DataFrame, or parquet format, e.g. >>> AfCycDesign.data.column('sequence') >>> AfCycDesign.to_pandas() >>> AfCycDesign.to_parquet("dataset.parquet") ### 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. ## 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. ## 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 ``` ## 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# 数据集卡片元数据参考规范：https://github.com/huggingface/hub-docs/blob/main/datasetcard.md?plain=1 # 文档/指南：https://huggingface.co/docs/hub/datasets-cards license_name: 知识共享署名4.0国际许可协议（Creative Commons Attribution 4.0 International） repo: https://doi.org/10.5281/zenodo.15164650 tags: - 环肽（cyclic peptides） - 蛋白质设计（protein design） - AlphaFold2 - 幻觉生成（hallucination） - 结构生物学（structural biology） pretty_name: AfCycDesign 幻觉生成支架 configs: - config_name: main data_files: - split: afcycpep_hallucinated path: data/afcycpep_data_hallucinated.csv - split: afcycpep_experimental path: data/afcycpep_data_experimental.csv # AfCycDesign 数据集卡片 本数据集为AfCycDesign用于环肽设计的幻觉生成支架。 ## 数据集详情 包含7至16聚体的幻觉生成肽CIF（Crystallographic Information File）文件与实验得到的CCDC结构。 ### 数据集描述 本数据集包含由AfCycDesign生成的幻觉环肽支架结构（格式为CIF）。AfCycDesign是一种基于AlphaFold2构建的深度学习方法，用于环肽的从头设计。该数据集的支架覆盖7至16个残基的肽链长度，通过同时对序列与结构进行采样的幻觉生成流程生成，可得到有序排列的环肽。数据集中的结构经AfCycDesign筛选，保留了预测置信度较高（pLDDT（Predicted Local Distance Difference Test）>0.9）的结构，共包含24104个结构多样的环肽，这些肽被预测可折叠为其设计的构象。这些支架可作为引入功能元件（如用于治疗应用的靶标结合基序）的起始结构。 - **整理者**：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 - **语言（自然语言处理）**：不适用（结构生物学数据） - **许可协议**：MIT许可证 ### 数据集来源 - **仓库**：https://zenodo.org/records/15164650 - **论文（可选）**：Rettie, S.A., Campbell, K.V., Bera, A.K. 等. 基于AlphaFold2的环肽结构预测与设计. *Nature Communications* **16**, 4730 (2025). https://doi.org/10.1038/s41467-025-59940-7 ## 快速使用指南 ### 安装 Hugging Face Datasets 库 可通过Hugging Face [datasets](https://huggingface.co/docs/datasets/index)库将每个子集加载至Python环境中。首先，通过命令行安装`datasets`库： $ pip install datasets 随后，在Python环境中导入datasets库： >>> import datasets ### 加载数据集 加载`RosettaCommons/AfCycDesign`数据集： >>> AfCycDesign = datasets.load_dataset('RosettaCommons/AfCycDesign') Downloading readme: 9.67kB [00:00, 3.57MB/s] Downloading data: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████| 2.94M/2.94M [00:00<00:00, 6.99MB/s] Downloading data: 100%|█████████████████████████████████████████████████████████████████████████████████████████████| 544/544 [00:00<00:00, 4.22kB/s] Generating afcycpep_hallucinated split: 100%|███████████████████████████████████████████████████████████████| 20656/20656 [00:00<00:00, 246500.54 examples/s] Generating afcycpep_experimental split: 100%|█████████████████████████████████████████████████████████████████████████| 8/8 [00:00<00:00, 3582.58 examples/s] 数据集将以`datasets.arrow_dataset.Dataset`格式加载： >>> AfCycDesign DatasetDict({ afcycpep_hallucinated: Dataset({ features: ['ID', 'sequence', 'nmer', 'rosetta_score', 'hf_path', 'Type'], num_rows: 20656 }) afcycpep_experimental: Dataset({ features: ['ID', 'sequence', 'nmer', 'rosetta_score', 'hf_path', 'Type'], num_rows: 8 }) }) 该数据集采用列式存储格式，可直接访问、转换为pandas.DataFrame或Parquet格式，例如： >>> AfCycDesign.data.column('sequence') >>> AfCycDesign.to_pandas() >>> AfCycDesign.to_parquet("dataset.parquet") ### 直接使用 这些幻觉生成支架可作为环肽结合剂设计流程中的起始结构。研究人员可将已知的结合基序（如蛋白质-蛋白质互作界面中的热点环）嫁接至这些支架上，并借助ProteinMPNN与Rosetta等工具对剩余序列进行重新设计，以得到针对治疗性蛋白质靶标的环肽结合剂。此外，这些支架还可用于基准测试环肽结构预测方法，或作为种子结构用于后续的计算设计工作。 ### 超出适用范围的使用 这些支架属于计算设计模型，并非全部经过实验验证。不得将其用作实验测定的结构。本数据集未经过充分的实验表征（包括结构验证、结合实验与稳定性测试），不得直接用于治疗用途。该支架仅由标准L-氨基酸组成，原生不支持非标准氨基酸设计，但可在后续进行事后替换。 ## 数据集结构 数据集中的每个条目均包含一个用于幻觉生成数据的PDB文件，该文件存储了幻觉生成环肽支架的三维原子坐标。支架的残基长度为7至16个，并按肽链大小分为7至16组。每个支架代表一个基于二面角分箱聚类得到的独特结构簇，其中每个残基根据其φ、ψ与ω主链二面角被分配至A、B、X或Y分箱中。本数据集的所有支架均经AfCycDesign预测，可高置信度折叠为其设计的结构（pLDDT>0.9）。实验条目则包含CIF文件。 ### 整理依据 生成这些支架的目的是提供一个大规模、结构多样且折叠良好的环肽骨架文库，用于后续的功能设计。传统的基于物理模型的方法（如Rosetta运动学闭合方法）计算成本高昂，且在不添加额外二硫键交联的情况下，难以处理较大的大环肽（11至13个残基）。AfCycDesign的幻觉生成流程克服了这些局限，可快速枚举多样的环肽拓扑结构，包括具有短α螺旋、β折叠与仅含环区构象的结构。 ### 源数据 https://zenodo.org/records/15164650 #### 数据收集与处理 支架通过AfCycDesign幻觉生成流程生成，该流程对AlphaFold2进行了修改，引入了自定义的环相对位置编码矩阵。针对每个肽链大小（最初为7至13个残基，后续扩展至16个），共生成了48000个幻觉生成模型。结构通过基于二面角分箱的方法进行聚类，并选取每个簇中置信度最高的成员（pLDDT>0.9）。幻觉生成流程通过优化预测置信度指标（pLDDT与PAE（Predicted Alignment Error））以及分子内接触数的损失函数进行引导。该流程基于ColabDesign v1.1.2框架实现。 #### 源数据生产者 本数据由华盛顿大学蛋白质设计研究所的研究人员以及哈佛大学的合作者通过计算生成。主要贡献者包括Stephen A. Rettie、Katelyn V. Campbell、Asim K. Bera、Sergey Ovchinnikov与Gaurav Bhardwaj。 #### 标注者 标注由AfCycDesign流程与AlphaFold2的置信度指标通过算法自动生成。 #### 个人与敏感信息 本数据集不包含任何个人、敏感或隐私信息。所有数据均为合成肽序列的计算生成结构模型。 ## 偏差、风险与局限性 这些支架属于计算预测结果，并非所有都能在实验条件下按照设计的方式折叠。仅对代表性子集进行了验证（8个X射线晶体结构，与设计模型的RMSD均小于1.0 Å），但完整数据集尚未经过实验表征。该幻觉生成流程仅支持20种标准L-氨基酸，原生不包含D-氨基酸或其他非标准残基。对于11个残基以上的肽链，48000次幻觉生成运行可能未完全覆盖可用的结构空间，意味着存在更多独特的支架。用于筛选的置信度指标（pLDDT）虽可较强地预测结构准确性，但无法保证实验成功。 ### 建议 用户应注意，这些支架大多为计算预测的结构，而非实验测定的结构。 ## 术语表 - **AfCycDesign**：本文中描述的深度学习框架，通过引入环相对位置编码，将AlphaFold2适配至环肽结构预测与设计任务。 - **pLDDT（Predicted Local Distance Difference Test）**：AlphaFold2输出的每个残基置信度指标，取值范围为0至1，数值越高代表预测置信度越高。 - **PAE（Predicted Alignment Error）**：AlphaFold2输出的指标，用于估计成对残基相对位置的误差。 - **Hallucination**：一种从头设计方法，从随机序列出发，通过优化AlphaFold2的置信度指标，同时生成序列与结构。 - **Pnear**：源自Rosetta的指标（取值0至1），表示折叠倾向；取值为1意味着设计的结构为单一最低能量构象。 - **CIF（Crystallographic Information File）**：用于表示三维分子结构的标准文件格式。 - **Torsion bin clustering**：一种通过将主链φ、ψ与ω二面角分配至分箱（A、B、X、Y）来对肽结构进行分组的方法。 ## 更多信息 AfCycDesign代码基于ColabDesign框架实现。如需了解方法与补充数据的更多细节，请参阅发表于https://doi.org/10.1038/s41467-025-59940-7的完整论文及其补充材料。 ## 引用格式 **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 ## 数据集卡片作者 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