MF-PCBA-Bind

Code for generating these datasets can be found at https://github.com/Leash-Labs/mf-pcba-bind

Protein-ligand binding prediction datasets derived from the MF-PCBA benchmark.

This repository extends the original MF-PCBA dataset by:

• Filtering to binding assays only (excluding phenotypic assays)
• Removing PAINS (pan-assay interference compounds) that show non-specific activity
• Providing pre-built validation and test splits for protein-ligand binding prediction
• Aggregating the data across assays into combined, easy-to-access files
• Providing "compact" versions of each val/test sets to be more accessible to expensive models

Attribution

This work is based on MF-PCBA by David Buterez et al. The original dataset, retrieval scripts, and methodology are preserved in the original-mf-pcba/ directory.

If you use this dataset, please cite the original MF-PCBA paper:

@article{doi:10.1021/acs.jcim.2c01569,
    author = {Buterez, David and Janet, Jon Paul and Kiddle, Steven J. and Li�, Pietro},
    title = {MF-PCBA: Multifidelity High-Throughput Screening Benchmarks for Drug Discovery and Machine Learning},
    journal = {Journal of Chemical Information and Modeling},
    year = {2023},
    doi = {10.1021/acs.jcim.2c01569},
    URL = {https://doi.org/10.1021/acs.jcim.2c01569}
}

Dataset Overview

Data Files

File	Description
data/mf_pcba_bind_val_full.parquet	Full validation set with all binders and non-binders
data/mf_pcba_bind_val_compact.parquet	Compact validation set (binders + 4x sampled non-binders)
data/mf_pcba_bind_test_full.parquet	Full test set with all binders and non-binders
data/mf_pcba_bind_test_compact.parquet	Compact test set (binders + 4x sampled non-binders)
data/mf_pcba_bind_val+test_full.parquet	Combined val+test full set
data/mf_pcba_bind_val+test_compact.parquet	Combined val+test compact set
data/MF-PCBA-Assay-Metadata.csv	Curated metadata for all assays including protein sequences

Columns

Each parquet file contains:

• CID: PubChem Compound ID
• smiles: Molecular SMILES string
• binds: Binary label (1 = binder, 0 = non-binder)
• protein_name: Target protein name
• protein_category: A rough categorization of the protein
• protein_accession: Protein accession number
• amino_acid_sequence: Full protein sequence
• AID: PubChem Assay ID

Label Definitions

• Binders (binds=1): Compounds marked "Active" in dose-response (DR) confirmatory screening
• Non-binders (binds=0): Compounds marked "Inactive" in single-dose (SD) primary screening

PAINS Filtering

PAINS (Pan-Assay Interference Compounds) are filtered out using RDKit's FilterCatalog. These are compounds that tend to show activity across many assays due to non-specific mechanisms (e.g., aggregation, redox cycling, fluorescence interference) rather than genuine target binding.

Scripts

scripts/build_val_test_sets.py

Builds the validation and test parquet files from retrieved MF-PCBA data:

1. Reads manually reviewed and curated assay metadata from data/MF-PCBA-Assay-Metadata.csv
2. Filters to binding assays only (excludes phenotypic assays)
3. Extracts binders (DR active) and non-binders (SD inactive)
4. Removes PAINS compounds using RDKit's FilterCatalog
5. Adds protein sequence information to each compound
6. Creates full and compact versions of validation, test, and combined sets

Requirements: pandas, numpy, pyarrow, rdkit

Usage:

# First, retrieve the raw MF-PCBA data using scripts in original-mf-pcba/
# Then run:
python scripts/build_val_test_sets.py

Original MF-PCBA

The original-mf-pcba/ directory contains the original MF-PCBA retrieval code and scripts. See original-mf-pcba/README-original.md for details on downloading and processing the raw PubChem data.

License

MIT License - see LICENSE

Original MF-PCBA code copyright (c) 2022 David Buterez.