MolPILE dataset

ArXiv preprint: "MolPILE - large-scale, diverse dataset for molecular representation learning" J. Adamczyk, J. Poziemski, F. Job, M. Król, M. Makowski

GitHub repository: https://github.com/scikit-fingerprints/MolPILE_dataset

Description

MolPILE is a large-scale molecular dataset designed for pretraining and evaluating machine learning models in cheminformatics. It is compiled from several major chemical databases: UniChem, PubChem, Mcule, ChemSpace, SuperNatural3, and COCONUT. Workflow included preprocessing, standardization, and feasibility filtering of molecules.

Data dictionary

Columns and meaning:

• source

  Source dataset identifier. Combined with id column, it provides full traceability to the original data entry.

  Possible values: PubChem, UniChem, Mcule, ChemSpace, SuperNatural3, COCONUT

• id

  A unique identifier for the molecule, retained from its source database (e.g., Mcule ID, PubChem CID).

  Example: MCULE-7212330550

• SMILES:

  Canonical SMILES representation of the molecule, after RDKit processing.

  Example: O=C(O)C1=CC=CC(O)C1O

Statistics

Workflow

Loading the MolPILE dataset

import polars as pl
from huggingface_hub import hf_hub_download

parquet_file_path = hf_hub_download(
    repo_id="scikit-fingerprints/MolPILE",
    repo_type="dataset",
    filename="molpile_smiles.parquet",
    local_dir="datasets",
    local_dir_use_symlinks=False
)

df = pl.read_parquet(parquet_file_path)
print(df.head())

Licenses

MolPILE is a collection of processed datasets, not a single dataset. It is shared as a single Parquet file only for convenience. Individual sources are separated by source and id columns. Each source has its own separate license, which we list below. As its entirety, MolPILE does not have a single license, as it is a collection, not a single dataset. Users interested in that can easily filter the dataset by source. In case of PubChem and UniChem, users may also want to check the individual licenses of their sources. Users using those sources are also asked to cite the appropriate publications, which we provide below. We do not make any claims about licensing of models trained on MolPILE, nor put any additional limitations.

• UniChem: CC0 (public domain); Chambers, Jon, et al. "UniChem: a unified chemical structure cross-referencing and identifier tracking system." Journal of Cheminformatics 5.1 (2013): 3.
• PubChem: CC-BY-4.0; Kim, Sunghwan, et al. "PubChem 2023 update." Nucleic Acids Research 51.D1 (2023): D1373-D1380.
• Mcule: CC-BY-NC-4.0; Kiss, Robert, Mark Sandor, and Ferenc A. Szalai. "http://Mcule.com: a public web service for drug discovery." Journal of Cheminformatics 4.Suppl 1 (2012): P17.
• ChemSpace: CC-BY-NC-4.0; please cite link https://chem-space.com/compounds/screening-compound-catalog
• SuperNatural3: not specified, only "freely available"; Gallo, Kathleen, et al. "SuperNatural 3.0 - a database of natural products and natural product-based derivatives." Nucleic Acids Research 51.D1 (2023): D654-D659.
• COCONUT: CC0 (public domain); Sorokina, Maria, et al. "COCONUT online: collection of open natural products database." Journal of Cheminformatics 13.1 (2021): 2.