manumartinm/bbb-peptides

BBB Peptide Dataset (TFG)

Curated blood–brain barrier (BBB) permeability dataset for peptide sequences with Boltz-predicted 3D structures and physicochemical descriptors.

Summary

Field	Value
Rows	825
With structure	825
BBB+	410
BBB−	415
Variant	full

Files

• peptides.parquet — one row per peptide: sequence, label, splits, physicochemical features, structure quality metrics, relative structure paths
• structures/<sequence_hash>/coords.npz — CA coordinates + per-residue pLDDT (coords, sequence, plddt_per_residue)
• structures/<sequence_hash>/structure.cif — Boltz predicted structure (when available)
• stats.json — export statistics

Columns (main)

• Identifiers: peptide_id, source_id, sequence, sequence_hash, length
• Label: bbb_label (1 = BBB-permeable, 0 = non-permeable)
• Splits: fold_id, cluster_id, external_test, split, source_split, is_gold
• Structure quality: plddt, ptm, structure_confidence, iptm, complex_iplddt, complex_pde, complex_ipde
• Structure paths: structure_coords_path, structure_cif_path, has_structure
• Physicochemical: mw, ext_coef_reduced, ext_coef_oxidized, hydrophobic_ratio_pct, pi, net_charge_ph7, total_charge, mean_hydrophobicity, hydrophobicity_ph7, hydrophilic_ratio, aliphatic_index, boman_index...

Load with Hugging Face Datasets

from datasets import load_dataset
import numpy as np
from pathlib import Path

repo_dir = "."  # clone of this dataset repo
ds = load_dataset("parquet", data_files=f"{repo_dir}/peptides.parquet", split="train")

row = ds[0]
coords = np.load(Path(repo_dir) / row["structure_coords_path"])
ca = coords["coords"]          # (L, 3) float32
seq = "".join(coords["sequence"])  # amino-acid string
plddt = coords["plddt_per_residue"]  # (L,) per-residue confidence

Load with pandas

import pandas as pd
df = pd.read_parquet("peptides.parquet")