license: cc-by-4.0
viewer: true
😊 Hi there. This is CPSea2, a data framework for structurally diverse cyclic peptide binder design.
🗺️ This is the site we provide our main data resources. We also provide the codes for dataset curation in the CPSea2 repo, and provide a usage demo of CPbase that a generative model, PepAny, can be trained on this dataset for linear peptide binder generation with controllable terminal amide geometries via CFG.
Based on the idea of cyclic peptides can be decomposed to a linear peptide and a capping structure, we constructed the two components separately, and combine them together based on the geometry of shared amides. We therefore provide two types of dataset for each.
CPbase
CPbase is a large linear peptide -- protein complex dataset capturing intra-protein interfaces to mimic inter-chain interactions. The overall curation idea of CPbase is rooted in CPSea, while key improvements in metrics and thresholds are involved and the scope is expanded to linear peptides. There are two source datasets for CPbase construction:
(1) For CPbase_PDB, we provide the structures and the metadata in CPbase_PDB/CPbase_PDB_Structures.tar.gz and CPbase_PDB/CPbase_PDB_metadata.tsv.gz. We also provide a demo subset containing 1000 .cif.gz files for you to give our dataset a galance, which is in CPbase_PDB_demo directory.
The metadata file contains basic informations of each ligand -- receptor pair. The format is:
<LigandID>\t<OriginalPosition>\t<Length>\t<absBSA>\t<relBSA>\t<raBSA>\t<BSA_per_length>\t<rel_n_rBSA>\t<rel_c_rBSA>\t<GRAVY>\t<HelixRatio>\t<SheetRatio>\t<SeparateChain>\t<LigandAllUpstream>\t<LigandAllDownstream>\t<InterchainDisulfide>\t<OriginalLigandChain>\t<OriginalReceptorChain>
Some notes:
- The length here does not include the capping residues.
SeparateChain,LigandAllUpstreamandLigandAllDownstreamare boolian values that annotate whether the receptor can be represented as a continuous sequence without any breakpoints, which may be compatible with protein folding models.InterchainDisulfideis also a boolian value that reports close interchain SG atoms that might be disulfide bonds. This information may be useful for scenerios such as covalent binder design.
The structure files are saved in the mmcif format, where the first structure paragraph is the residue union of receptors, and the following structures are different peptide ligands. We use the _cpsea_receptor_ligand_map to record the mapping between each receptor residue and its corresponding ligand ID, so that the receptor of a ligand can be easily retrieved.
(2) For CPbase_AFDB, although we follow a similar protocol and have collected the structures, the size of structure files is too large to be properly handled in data host sites. Nonetheless, because we only used the AFDB monomer data and no post-processing has been performed, a complex can be fully described by res_ids. Therefore, we alternatively provide the res_id maps in CPbase_AFDB/ResID_Map. Please contact us if you would like to access the strcutrue files. We also provide full metadata files in CPbase_AFDB/Metadata, with the same format as described above, without the last two columns for the original chainID because all residues in AFDB have a chainID of A.
Similarly, a demo of CPbase_AFDB is available in CPbase_AFDB_demo, where we provide the metadata, res_id and structure files to enable a preview.
CPcap
CPcap is generated by simple MM conformer sampling as described in our repo. We currently provide a demo of 8 types of cyclization structures, while we believe that more diverse structures can be readily involved using a similar protocol.
In the CPcap directory, we provide the conformer ensemble of the 8 cyclization types, as well as their energy annotations provided by Schrodinger.
Merged CPcap and CPbase
With both CPcap and CPbase in hand, we can construct cyclic peptide -- protein complex structures by aligning the two dataset based on the RMSD of their shared amides. We provide the raw merged structures and the relaxed structures in Merged_CPcap_CPbase, devided to different directories based on cyclization types. We also provide metadata files annotating the source of the merged structure.