PubChemQC PM6: Data Sets of 221 Million Molecules with Optimized Molecular Geometries and Electronic Properties

Nakata Maho, Maho Nakata, Tomomi Shimazaki, Masatomo Hashimoto, Toshiyuki MaedaJournal of Chemical Information and Modeling 2020 · DOI

TL;DR

Researchers have created the largest publicly available dataset of molecular geometries and electronic properties, PubChemQC PM6, by performing quantum chemical calculations on 221 million molecules. This massive dataset, generated using the PM6 semi-empirical method, aims to accelerate machine learning research in chemistry by providing a rich source of high-quality training data for predicting molecular properties.

What Problem Does This Paper Address?

The rapid advancement of organic chemistry, particularly in areas like materials science and drug discovery, necessitates the exploration of an ever-expanding chemical space. Traditional methods of discovering and characterizing new molecules, such as physical experiments, are time-consuming and expensive. While ab initio quantum chemical calculations offer accurate predictions, their computational cost is prohibitive for exploring the vastness of chemical space (estimated at 10^60 drug candidates alone). Machine learning has emerged as a powerful tool for accelerating this process, but it relies heavily on large, high-quality datasets for training. Currently, such comprehensive datasets for molecular properties are scarce, posing a bottleneck for developing predictive models. This paper addresses this gap by generating a massive, publicly accessible dataset of molecular properties.

Methods

The core of this work involves performing quantum chemical calculations on a large number of molecules. Specifically, the authors utilized the PM6 semi-empirical method for geometry optimization and electronic structure calculations. PM6 is chosen for its balance of speed and accuracy, making it suitable for large-scale computations.

Key methodological aspects include:

  • Geometry Optimization: The PM6 method was used to find the lowest energy conformation (optimized geometry) for each molecule.
  • Electronic Property Calculation: After geometry optimization, electronic properties were calculated.
  • Multiple Electronic States: Calculations were performed not only for neutral states but also for cationic, anionic, and spin-flipped states to provide a more comprehensive property set.
  • Molecular Representation: The authors leveraged SMILES (Simplified Molecular Input Line Entry System) for generating initial geometry guesses and InChI (International Chemical Identifier) for validating the optimized results. These standardized string representations are crucial for handling and identifying molecules programmatically.
  • Computational Infrastructure: The calculations were performed on high-performance computing clusters (RIKEN HOKUSAI BigWave, QUEST, and RICC), highlighting the scale of the undertaking.
  • Data Processing Pipeline: A robust pipeline was developed to parse molecular data, perform calculations, and store the results, including checks for InChI validity to ensure data integrity.

Datasets

The primary dataset used in this study is PubChem Compounds. This is a publicly available database maintained by the National Institutes of Health (NIH), containing a vast collection of unique chemical compounds.

The authors generated a new dataset, PubChemQC PM6, which is the main contribution of this paper. This dataset comprises:

  • 221 million molecules in total.
  • Calculations for 92.9% of the 91.2 million molecules cataloged in PubChem Compounds as of August 29, 2016.
  • Includes calculations for neutral states, as well as cationic (56.2%), anionic (49.7%), and spin-flipped (41.3%) electronic states.
  • The dataset is available at http://pubchemqc.riken.jp/pm6_dataset.html under a Creative Commons Attribution 4.0 International License.

Key Insights and Ideas

  • The sheer scale of chemical space: The paper underscores the immense number of possible molecules, emphasizing the need for computational tools like machine learning to navigate it.
  • The trade-off between speed and accuracy: The choice of PM6 highlights the practical necessity of balancing computational cost with the desired level of accuracy for large-scale datasets. While DFT methods are more accurate, PM6 offers a significant speed advantage for generating a dataset of this magnitude.
  • The importance of standardized molecular representations: The reliance on SMILES and InChI demonstrates their critical role in managing and validating large chemical datasets, enabling reproducibility and downstream analysis.
  • Challenges in data curation: The paper details numerous issues encountered with the PubChem database, including duplicates, inconsistent InChI/SMILES representations, and molecules that are difficult to optimize. This highlights the ongoing challenges in creating and maintaining high-quality chemical data.
  • The potential for machine learning: By providing such a large and diverse dataset, the authors aim to significantly boost the development of machine learning models for predicting molecular properties, accelerating discovery in various chemical fields.

Key Results

  • Dataset Size: A total of 221,190,415 molecules were calculated, comprising optimized geometries and electronic properties.
  • Coverage: The PubChemQC PM6 dataset covers 94.0% of neutral molecules in the PubChem Compound database.
  • Calculation Success Rate: Geometry optimization was successful for 86,213,135 neutral molecules.
  • InChI Validity: For 92.9% of the compounds, the original and calculated InChIs coincided up to the chemical formula and atom connection of the main layer, indicating good data integrity.
  • Computational Time: The entire calculation process took approximately 95 days on the HOKUSAI BigWave supercomputer, demonstrating the efficiency of the PM6 method for large-scale computations.
  • Failure Analysis: The paper provides detailed statistics on calculation failures, particularly noting difficulties with lanthanoids, actinoids, and many metals, which will inform future research.

Limitations and Future Work

  • Limitations:
    • PM6 Accuracy: While PM6 offers speed, it is a semi-empirical method and thus less accurate than ab initio methods like DFT. The authors acknowledge that DFT methods provide better results but are computationally prohibitive for this scale.
  • Isotopes and Charged Ions: Isotopes and charged ions were not considered in the calculations.
  • Data Ambiguities: The paper discusses issues with duplicates, inconsistent molecular representations (SMILES/InChI), and molecules that are inherently unstable or difficult to optimize, leading to some data inconsistencies.
  • Solids and Solutions: The current approach is primarily for gas-phase molecules, and the methodology does not directly apply to solids or solutions.
  • Future Work:
    • Detailed Analysis: The authors plan to publish more detailed analyses of properties like HOMO-LUMO gaps, energies, vibration intensities, and dipole moments.
  • Database Dump: They intend to provide a database dump for easier access and utilization by researchers.
  • B3LYP Calculations: They are running B3LYP calculations using the PM6-optimized geometries to provide higher-accuracy data for a subset of molecules.
  • Machine Learning Models: The team is developing machine learning models for quantitative structure-property relationships (QSPR) based on this dataset.
  • Curation and Improvement: Further investigation into failed calculations and curation of successfully calculated molecules is planned, potentially leading to improvements in molecular encoding methods.

Summary updated

BibTeX
@article{maho2020pubchemqc,
  author = {NAKATA Maho and Maho Nakata and Tomomi Shimazaki and Masatomo Hashimoto and Toshiyuki Maeda},
  title = {{PubChemQC PM6: Data Sets of 221 Million Molecules with Optimized Molecular Geometries and Electronic Properties}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2020},
  doi = {10.1021/acs.jcim.0c00740},
}
Figure