CycPepMPDB

CycPeptMPDB (Cyclic Peptide Membrane Permeability Database) is the standard public dataset of experimentally measured membrane permeability for cyclic peptides, released in 2023 by Akiyama and colleagues at Tokyo Institute of Technology. It exists because cyclic peptides have become a serious modality for hitting "undruggable" intracellular targets like protein-protein interfaces, but their typically poor passive permeability is the bottleneck of the field, and until CycPeptMPDB there was no unified resource to train or benchmark in silico permeability predictors. The database aggregates 7,334 cyclic peptides curated from 45 published papers and 2 pharma patents, with experimentally measured permeability values (predominantly PAMPA and Caco-2, also RRCK and MDCK depending on source) under harmonized units. Each entry carries the 2D structure, HELM sequence representation (which cleanly handles the >99% of entries containing non-natural building blocks like N-methylated residues, D-amino acids, and exotic backbones), and 3D conformer data, so the corpus is directly usable for chemoinformatic featurization, graph neural networks, conformational ensemble methods, and language-model approaches over HELM or SMILES. The website at cycpeptmpdb.com adds online visualization, filtering by source assay and scaffold, and bulk downloads. CycPeptMPDB has since become the backbone of essentially all recent cyclic peptide permeability ML work (CycPeptMP and its descendants).

Splits

• train: 7,983 rows
• test: 868 rows

Rows are assigned with a deterministic hash split: sha256(record_id) % 10, where bucket 0 is test and buckets 1-9 are train.

Columns

The table includes common browsing and modeling columns:

• record_id: stable SHA-256 row identifier
• dataset_id, source_file, source_table, source_row_index: source provenance
• entity_type: peptide or monomer
• assay_name, sequence, sequence_length, target, score_value, label: normalized convenience fields when present
• split_bucket: deterministic split bucket

Original descriptor and assay fields are preserved as snake_case string columns, including fields such as smiles, permeability, caco2, pampa, mdck, rrck, and molecular descriptor columns. See metadata/column_mapping.parquet for the original field-name mapping and metadata/source_tables.parquet for per-table row counts.

Usage

from datasets import load_dataset

ds = load_dataset("LiteFold/CycPepMPDB")
print(ds)
print(ds["train"][0])

Load selected columns:

from datasets import load_dataset

cols = ["record_id", "entity_type", "sequence", "smiles", "permeability", "score_value"]
train = load_dataset("LiteFold/CycPepMPDB", split="train", columns=cols)

Filter to peptides:

from datasets import load_dataset

train = load_dataset("LiteFold/CycPepMPDB", split="train")
peptides = train.filter(lambda row: row["entity_type"] == "peptide")

Metadata tables can be loaded directly:

from datasets import load_dataset

source_tables = load_dataset(
    "parquet",
    data_files="hf://datasets/LiteFold/CycPepMPDB/metadata/source_tables.parquet",
    split="train",
)
column_mapping = load_dataset(
    "parquet",
    data_files="hf://datasets/LiteFold/CycPepMPDB/metadata/column_mapping.parquet",
    split="train",
)

Citation

@article{li2023cycpeptmpdb,
  title     = {{CycPeptMPDB}: A Comprehensive Database of Membrane Permeability of Cyclic Peptides},
  author    = {Li, Jianan and Yanagisawa, Keisuke and Sugita, Masatake and Fujie, Takuya and Ohue, Masahito and Akiyama, Yutaka},
  journal   = {Journal of Chemical Information and Modeling},
  volume    = {63},
  number    = {7},
  pages     = {2240--2250},
  year      = {2023},
  publisher = {American Chemical Society},
  doi       = {10.1021/acs.jcim.2c01573}
}