Source-linked AI summary
QM7-X: A comprehensive dataset of quantum-mechanical properties spanning the chemical space of small organic molecules
Johannes Hoja, Leonardo Medrano Sandonas, Brian G. Ernst, Alvaro Vazquez-Mayagoitia, Robert A. DiStasio Jr., Alexandre Tkatchenko
TL;DR
QM7-X addresses the need for extensive quantum-mechanical data to study structure–property relationships across chemical compound space. It systematically samples small-molecule equilibrium and non-equilibrium structures and computes diverse properties; the resulting dataset covers 96.4% of GDB13 molecules with up to seven heavy atoms.
Problem
Comprehensive QM datasets are needed to study molecular structure–property relationships, but their size grows exponentially with molecule size and tightly converged calculations are computationally costly.
Source: Background & Summary · p. 1Method
QM7-X systematically samples isomers and conformers for molecules with up to seven heavy atoms, generates non-equilibrium structures, and computes 42 properties primarily at PBE0+MBD.
Source: Background & Summary · pp. 2–3Results
96.4% of GDB13 molecules with up to seven heavy atoms are covered by QM7-X, alongside 98.8% of their corresponding metastable stereoisomers.
Source: Generation of non-equilibrium structures · p. 4Takeaways & Limitations
The dataset provides stereoisomer-rich data that the authors expect will help ML models capture physicochemical differences relevant to drug design.
Source: Technical Validation · p. 6PBE0+MBD relative energies do not strictly follow the DFTB3+MBD Boltzmann distribution used to generate structures, although the distributions are often similar.
Source: Technical Validation · p. 7
Abstract
from arXiv · showhide
We introduce QM7-X, a comprehensive dataset of 42 physicochemical properties for $\approx$ 4.2 M equilibrium and non-equilibrium structures of small organic molecules with up to seven non-hydrogen (C, N, O, S, Cl) atoms. To span this fundamentally important region of chemical compound space (CCS), QM7-X includes an exhaustive sampling of (meta-)stable equilibrium structures - comprised of constitutional/structural isomers and stereoisomers, e.g., enantiomers and diastereomers (including cis-/trans- and conformational isomers) - as well as 100 non-equilibrium structural variations thereof to reach a total of $\approx$ 4.2 M molecular structures. Computed at the tightly converged quantum-mechanical PBE0+MBD level of theory, QM7-X contains global (molecular) and local (atom-in-a-molecule) properties ranging from ground state quantities (such as atomization energies and dipole moments) to response quantities (such as polarizability tensors and dispersion coefficients). By providing a systematic, extensive, and tightly-converged dataset of quantum-mechanically computed physicochemical properties, we expect that QM7-X will play a critical role in the development of next-generation machine-learning based models for exploring greater swaths of CCS and performing in silico design of molecules with targeted properties.
Background & Summary
QM7-X addresses the need for comprehensive quantum-mechanical datasets by systematically sampling small-molecule structures and providing tightly converged properties across chemical compound space. It combines equilibrium and non-equilibrium structures with global and local properties computed primarily at PBE0+MBD.
Background & Summary
- Earlier GDB datasets encode molecular formula and connectivity but lack structural or property information, while QM7 and QM9 contain one metastable structure per SMILES string.ANI-1 and ANI-1x provide larger collections of conformations and properties, but the paper identifies remaining gaps in systematic stereoisomer sampling, accuracy assessment, property breadth, and interaction modeling.Source: Background & Summary · pp. 1–2
- QM7-X targets four gaps in small-molecule datasets: systematic isomer and stereoisomer sampling, assessment of QM accuracy, broad local and global properties, and data for modeling covalent and van der Waals interactions.The need for comprehensive QM-based data is acute because dataset dimension grows exponentially with molecule size and tightly converged calculations are computationally costly.Source: Background & Summary · pp. 1–2
- The dataset systematically samples equilibrium structures for molecules with up to seven heavy atoms, then adds 100 normal-mode-displaced non-equilibrium structures per equilibrium structure, yielding approximately 4.2 million structures.Sampling includes constitutional and structural isomers, stereoisomers, and conformational isomers.Source: Background & Summary · p. 2, Figure 1 · p. 2
- QM7-X provides 42 molecular and atom-in-a-molecule properties, including ground-state quantities and response properties, computed mostly with PBE0+MBD and tightly converged numeric atom-centered orbitals.Examples include atomization energies, forces, dipole moments, polarizability tensors, and dispersion coefficients.Source: Background & Summary · p. 3
Methods
QM7-X structures were generated by enumerating molecular isomers and conformers, refining equilibrium geometries, and sampling non-equilibrium displacements. PBE0+MBD calculations then supplied the dataset's physicochemical properties.
Generation of equilibrium structures
- Canonical SMILES and MMFF94 conformer searches generated constitutional, structural, stereochemical, and conformational variants; DFTB3+MBD optimization then selected distinct conformers by RMSD.MMFF94 conformers were retained within 50 kcal/mol of the lowest-energy structure and at least 0.5 Å RMSD apart; DFTB3+MBD conformers had to exceed 1.0 Å RMSD from accepted structures.Source: Generation of equilibrium structures · p. 3
Generation of non-equilibrium structures
- Each equilibrium structure received 100 non-equilibrium structures sampled through DFTB3+MBD normal-mode displacements, with candidate energies checked against a Boltzmann distribution.The sampling used T = 1500 K and required candidate structures to fill a predefined energy histogram; 261 of 7,211 formulae were excluded when a complete distribution could not be obtained, leaving 96.4% of GDB13 molecules covered.Source: Generation of non-equilibrium structures · pp. 3–4
Calculation of physicochemical properties
- PBE0+MBD single-point calculations with tight settings provided energies, forces, and other properties for the DFTB3+MBD structures.MBD energies, atomic forces, atomic C6 coefficients, and isotropic atomic polarizabilities used rsSCS; molecular C6 coefficients and polarizabilities used SCS.Source: Calculation of physicochemical… · p. 4, Table 2 · p. 5
Data Records
QM7-X is distributed as HDF5 files with each structure's properties stored in a Python dictionary. Structure labels encode molecular identity, stereoisomer, conformer, and whether a geometry is optimized or displaced.
Data Records
- The dataset is distributed in eight HDF5 files in a Zenodo repository, with a README providing technical usage details and access examples.Each molecular structure's information is stored as a Python dictionary.Source: Data Records · p. 6, HDF5 file format · p. 6
HDF5 file format
- Labels of the form Geom-mr-is-ct-u identify the SMILES string, stereoisomer, conformer, and optimized or displaced structure; indices follow generation order, not energy order.The label suffix u = opt denotes an optimized geometry, while u = 1,...,100 denotes a displaced structure.Source: HDF5 file format · p. 6
Technical Validation
QM7-X samples stereoisomers and distorted structures around equilibrium conformers, while its composition and property distributions reveal both broad coverage and uneven representation. The authors also note that structure generation and property calculations use different levels of theory, with only minor differences in the resulting energy distributions.
Technical Validation
- QM7-X emphasizes stereoisomers and distinct metastable rotational conformers, giving flexible molecules with stereocenters more data than rigid molecules without them.The authors expect this coverage to help ML models capture physicochemical differences between stereoisomers.Source: Technical Validation · p. 6
- Each metastable equilibrium conformer is paired with 100 non-equilibrium structures, generated by normal-mode displacements and selected to sample surrounding potential-energy surfaces.Candidate structures were re-evaluated and retained to follow a Boltzmann energy distribution, avoiding reliance on the harmonic approximation's accuracy.Source: Technical Validation · p. 6
- DFTB3+MBD generated structures, while PBE0+MBD supplied subsequent energies, forces, and properties; their energy distributions are often similar and yield virtually the same average energy difference.The authors report minor shifts in the first histogram bins and an insignificant number of structures slightly outside the intended energy window.Source: Technical Validation · p. 7
- Seven-heavy-atom molecules are most abundant, with 37,438 equilibrium and 3,743,800 non-equilibrium structures; C5H11NO is the most plentiful formula, with 3,200 equilibrium structures.The composition schematic counts structures with at least two or three different heavy elements; all molecules contain C and H.Source: Figure 2 · p. 6, Technical Validation · p. 8
- The dataset contains 68.0% [H,C,N] structures, 65.0% [H,C,O], and 38.7% [H,N,O], while only 0.06% contain Cl.It also includes 53,833 [H,C,S], 33,936 [H,N,S], and 32,017 [H,O,S] structures.Source: Technical Validation · p. 8
- Structural distortions produce significant property fluctuations; intensive properties are more sensitive than extensive properties, and molecular polarizabilities and C6 coefficients show similar distributions.The latter similarity is attributed to their correlation via the Casimir-Polder integral; local analogs show peaks associated with specific atomic environments.Source: Figure 3 · p. 7, Technical Validation · p. 8
Competing interests
The authors declare no competing financial interests and cite their dataset in Zenodo.
Competing interests
- The authors declare no competing financial interests.Source: Competing interests · p. 8
Data Citation
- The dataset citation names Hoja and coauthors and identifies Zenodo record 3905361.Source: Data Citation · p. 8