Roche PDE10A Binding-Affinity Benchmark

A high-quality industrial benchmark for protein-ligand binding-affinity prediction, released by Roche from a legacy phosphodiesterase 10A (PDE10A) lead-optimization program. The activity table holds 1,162 PDE10A inhibitors with measured binding affinities (pIC50 from 4.76 to 10.10), canonical SMILES, an assigned binding-mode class spanning three chemical series, and the assay and structure-solved dates. Alongside it are 1,162 docked protein-ligand complexes in PDBbind format, generated against the program's 77 PDE10A X-ray co-crystal structures. The benchmark ships with seven predefined evaluation splits: a random split, three temporal splits (2011, 2012, 2013) that partition compounds by measurement date, and three series splits that hold out an entire chemical class as the test set. The temporal and series splits measure extrapolation to new chemotypes, the regime that matters during real lead optimization, where the original study found that machine-learning models degrade sharply relative to their random-split performance. Published by Tosstorff et al., Journal of Computer-Aided Molecular Design (2022). The structural data is redistributed from the authors' figshare deposit (10.6084/m9.figshare.20055014) under the MIT license; the activity table is the paper's supplementary data. Hosted by Scigantic for direct mounting.

Category: Drug Discovery
Provider: Tosstorff et al. — Roche Pharma Research & Early Development
License: MIT
DOI: 10.1007/s10822-022-00478-x
Hosted on: Amazon S3, bucket scigantic-pde10a-benchmark, us-east-1
Access: Free and public. No account with the data provider is needed.
Source: https://link.springer.com/article/10.1007/s10822-022-00478-x

Data structure and loading code

Fields

`compound_id`	int — row index 0..1161
`SMILES`	string — canonical SMILES of the inhibit
`binding_mode_class`	string — one of 3 chemical series: amino
`pic50`	float — -log10(IC50 in molar); 4.76..10.
`pic50_date`	date (DD-Mon-YYYY) — assay readout date;
`DATESOLVED`	date (DD-Mon-YY) — date the X-ray struct
`docking_folder`	string — join key into structures/.../<d
`random_split`	string — train(697)/val(232)/test(233);
`temporal_2011_split`	string — train/val/test by 2011 cutoff;
`temporal_2012_split`	string — train/val/test by 2012 cutoff;
`temporal_2013_split`	string — train/val/test by 2013 cutoff;
`aminohetaryl_c1_amide_split`	string — series holdout: the 452 aminohe
`aryl_c1_amide_c2_hetaryl_split`	string — series holdout: the 419 aryl_c1
`c1_hetaryl_alkyl_c2_hetaryl_split`	string — series holdout: the 291 c1_heta

Load the data

# PDE10A binding-affinity benchmark — start from the activity table.
import os, pandas as pd
MNT = os.environ.get('SCIGANTIC_MOUNT_PATH', '/mnt/archive')  # also symlinked at ~/archive
df = pd.read_csv(f'{MNT}/pde10a_inhibitors_activity_splits.csv')
print(df.shape)                       # (1162, 14)
print(df['pic50'].describe())         # target: -log10(IC50 in M), 4.76..10.10
print(df['binding_mode_class'].value_counts())

# Use a HARD split (temporal 2012), not random, for an honest prospective test:
s = 'temporal_2012_split'
train = df[df[s] == 'train']; val = df[df[s] == 'val']; test = df[df[s] == 'test']
print({k: len(v) for k, v in {'train': train, 'val': val, 'test': test}.items()})
# X = train['SMILES'] -> featurize (RDKit / a SMILES model); y = train['pic50']

# Structures are pre-extracted (no unzip): join via docking_folder.
folder = df.iloc[0]['docking_folder']  # e.g. 5sf8_0
base = f'{MNT}/structures/pde-10_pdb_bind_format_blinded/{folder}'
print(os.listdir(base))  # <id>_protein/.pocket/.ligand in .pdb/.mol2/.sdf

Written for a Scigantic notebook, where /mnt/archive is the mounted dataset.

Analyze this dataset without downloading it

On Scigantic, Roche PDE10A Binding-Affinity Benchmark opens in a hosted JupyterLab notebook with the data already mounted, so you can read files where they are stored instead of copying them first. Each notebook includes an AI assistant that can see the mounted files and write and run analysis code. Open the dataset page to launch a notebook, or browse the full catalog.

Roche PDE10A Binding-Affinity Benchmark

Fields

Load the data

Analyze this dataset without downloading it

Related Drug Discovery datasets