TL;DR
Researchers have released the PubChemQC B3LYP/6-31G*//PM6 dataset, a massive open-source collection of electronic properties for nearly 86 million molecules. By providing high-level quantum chemical data for 94% of the PubChem catalog, this resource offers the scale and diversity needed to train the next generation of machine learning models for drug discovery and materials science.
What Problem Does This Paper Address?
In the search for new drugs or high-performance materials (like OLEDs or battery electrolytes), researchers need to know a molecule's electronic properties—such as its HOMO-LUMO energy gap or dipole moment. While experimental measurements are expensive and slow, quantum chemical (QC) calculations offer a reliable alternative.
However, high-accuracy QC methods like Density Functional Theory (DFT) are computationally demanding. Existing datasets are either too small (like QM9, with ~134,000 molecules) or lack chemical diversity. This "data gap" prevents machine learning models from generalizing well to the vast "chemical space" of possible molecules. The authors aim to bridge this gap by providing a dataset that is orders of magnitude larger and more diverse than anything previously available.
Methods
The authors employed a "best of both worlds" computational strategy:
- Geometry Optimization: They used the PM6 method, a fast semi-empirical approach, to find the stable 3D shapes of 86,213,135 neutral molecules.
- Electronic Structure: On these optimized shapes, they performed single-point calculations using B3LYP/6-31G, a widely respected "workhorse" DFT method. This combination is denoted as B3LYP/6-31G//PM6**.
- Heavy Lifting: The project required 31.1 million core hours on the RIKEN HOKUSAI BigWave supercomputer and the QUEST cluster, resulting in over 50 TB of data.
- Deployment: To make the data useful, they provided it in three formats: raw GAMESS files, structured JSON, and a Dockerized PostgreSQL database for easy querying.
Datasets
The primary source is the PubChem Compound catalog (as of August 2016). From this, the authors generated:
- The Full Set: 85,938,443 molecules (after removing failed calculations).
- Five Targeted Subsets: These are categorized by atom type (CHON vs. CHNOPSFCl) and molecular weight (under 300 or 500 Da) to help researchers focus on specific chemical domains, such as small drug-like molecules or essential biomolecules.
This dataset is roughly 130 times larger than other recent large-scale efforts like QMugs.
Key Insights and Ideas
- The "Good Enough" Geometry: A major finding is that PM6-optimized geometries are remarkably close to those optimized with more expensive DFT methods (median bond length difference of only 0.016 Å). This justifies using PM6 for the "shape" and DFT for the "electronics" at this massive scale.
- Strong Correlations: The authors found high coefficients of determination ($R^2$ up to 0.945) when comparing properties calculated via PM6 vs. B3LYP. This suggests that even if a researcher only has access to fast semi-empirical methods, the resulting data is highly structured and predictable.
- Beyond SMILES: While many ML models rely on 1D SMILES strings, the authors emphasize that 3D atomic coordinates are fundamental for reaching "chemical accuracy."
- Infrastructure as a Service: By providing a Dockerized PostgreSQL setup, the authors have lowered the barrier to entry for data scientists who may not be experts in quantum chemistry software but need to query millions of molecular properties.
Key Results
- Scale: 85.9 million molecules covering 94% of PubChem.
- Accuracy Benchmarks: The dataset shows strong internal consistency. For the HOMO-LUMO gap, $R^2$ values ranged from 0.803 to 0.892 across different subsets.
- ML Potential: The authors note that recent models (like GPS++) trained on similar data have achieved a Mean Absolute Error (MAE) of 0.072 eV for energy gap estimation—approaching the "gold standard" of chemical accuracy (0.043 eV).
Limitations and Future Work
- The Conformer Problem: The authors used initial 3D guesses from Open Babel. They acknowledge they did not perform exhaustive conformer searches, meaning some molecules might not be in their absolute lowest-energy state.
- Level of Theory: While B3LYP/6-31G* is a standard, it is not the most advanced functional available today.
- Future Directions: The team plans to use the CREST program for better conformer sampling and move toward even higher-level theories (like ωB97X-D) with improved basis sets. They also aim to develop a web interface for easier data visualization.
Summary updated