CPSea

🚧 We highly recommend using the new version of CPSea which we will release soon! Current version is with some issues about hydrogen geometry and chirality!

😊 Hi, this is CPSea, a large scale cyclic peptide protein complex database. Containing 2.64M complex structures, CPSea enables training cyclic peptide design models from scratch for the first time. We hope that everyone can find helpful parts from CPSea 🌊

🗺️ This site and Kaggle site store the main body of CPsea, as well as some list files and metadata for different subsets. Please check our GitHub for relevant scripts for generation and evaluation of CPSea.

Dataset

💡 Below is the structure of the dataset:

.
├── CPBind/
│   ├── CPBind_index.txt
│   ├── CPBind_pdb/
│   │   ├── CPBind_pdb.1
│   │   ├── CPBind_pdb.2
│   │   ├── CPBind_pdb.3
│   │   └── CPBind_pdb.4
│   └── CPBind_Properties/
│       ├── CPBind_affinity.csv
│       ├── CPBind_basic.tsv
│       ├── CPBind_cluster.tsv
│       ├── CPBind_hydrophobic.csv
│       └── CPBind_validity.csv
├── CPBind_unique/
│   ├── CPBind_unique_index.txt
│   ├── CPBind_unique_pdb
│   └── CPBind_unique_Properties/
│       ├── CPBind_unique_affinity.csv
│       ├── CPBind_unique_basic.tsv
│       ├── CPBind_unique_hydrophobic.csv
│       └── CPBind_unique_validity.csv
├── CPSea_AFDB/...
├── CPSea_AFDB_unique/...
├── CPTrans/...
├── CPTrans_unique/...
├── CPCore/...
├── CPCore_unique/...
├── CPSea_PDB/...
├── CPSea_PDB_unique/...
└── Weight/
    ├── DiffPepBuilder_Cyc.pth
    ├── PepFlow_Cyc.pt
    └── PepGlad_Cyc.ckpt

We provide 10 datasets, including CPSea derived from AFDB and PDB, and 3 subsets of CPSea_AFDB. These 5 datasets are clustered based on Foldseek, and we select one complex from every cluster, to create corresponding _unique datasets. Basic information of the provided datasets are listed below:

In each dataset, we provide an index file, structure tar file(s), and property files. 

• In index file, a list of file names is provided, which corresponds to files in the structure tar file. These names are in a format of {original_structure_id}_{first_residue}_{last_residue}_relaxed. Therefore, the original structure from which each cyclic peptide is derived can be known from the file.
• For structure tar file(s), we use _pdb suffix as identifier (e.g., CPBind_unique_pdb). The .tar extension is removed, and large tar files are chunked into several files for efficient processing in Kaggle (e.g. CPSea_AFDB_pdb.1, CPSea_AFDB_pdb.2, etc.). Though there are no file extension, these files are tar files, and you may run tar -xvf <structure_tar_file> to extract pdb files.
• For the property files, we provide 5 kinds of csv/tsv files, as described below:

Weights

We provide checkpoint file for the three models re-trained on CPSea in Weights, namely: DiffPepBuilder_Cyc.pth, PepFlow_Cyc.pt and PepGlad_Cyc.ckpt.

Please check out our GitHub for instructions on how to generate and evaluate cyclic peptides using these re-trained weights.