Source-linked AI summary
QMugs: Quantum Mechanical Properties of Drug-like Molecules
Clemens Isert, Kenneth Atz, José Jiménez-Luna, Gisbert Schneider
TL;DR
Curated quantum-chemical datasets for bioactive molecules remain scarce, limiting coverage of molecular space and cross-level learning. QMugs constructs a large ChEMBL-derived collection with optimized conformers, properties at GFN2-xTB and DFT levels, and wave functions. The resulting resource is intended to support molecular-property, conformational, and quantum-feature modeling in pharmacologically relevant data.
Problem
Curated datasets with broad chemical coverage and quantum-chemical information on bioactive molecules are limited, and such datasets for delta-learning are scarce.
Source: Background & Summary · p. 1Method
QMugs curates ChEMBL molecules, generates and optimizes conformers with GFN2-xTB, and computes properties and wave functions at GFN2-xTB and ωB97X-D/def2-SVP levels.
Source: Background & Summary · pp. 1–2Results
QMugs contains more than 665k molecules and about 2.0M conformers, including quantum properties and wave functions totaling over 7 terabytes of data.
Source: Background & Summary · pp. 1–2Takeaways & Limitations
The dataset is intended to support direct quantum-property prediction, cross-level learning, conformational modeling, wave-function prediction, and QML with annotated biological data.
Source: Background & Summary · p. 2
Abstract
from arXiv · showhide
Machine learning approaches in drug discovery, as well as in other areas of the chemical sciences, benefit from curated datasets of physical molecular properties. However, there is a lack of sufficiently large data collections that include first-principle quantum chemical information on bioactive molecules, such as single-point electronic properties, quantum mechanical wave functions and density-functional theory (DFT) matrices. The open-access QMugs (Quantum-Mechanical Properties of Drug-like Molecules) dataset fills this void. The QMugs collection comprises quantum mechanical properties of more than 665k biologically and pharmacologically relevant molecules extracted from the ChEMBL database, totaling $\sim$2M conformers. QMugs contains optimized molecular geometries and thermodynamic data obtained via the semi-empirical method GFN2-xTB. Atomic and molecular properties (e.g., partial charges, energies, and rotational constants) are provided on both the GFN2-xTB and on the DFT ($\omega$B97X-D/def2-SVP) levels of theory. QMugs also comprises the respective quantum mechanical wave functions, including DFT density and orbital matrices, totaling over 7 terabytes of uncompressed data. This dataset is intended to facilitate the development of models that learn from molecular data on different levels of theory while also providing insight into the corresponding relationships between molecular structure and biological activity.
Background & Summary
QMugs pairs ChEMBL-derived drug-like molecules with quantum properties computed at semi-empirical and DFT levels, including quantum mechanical wave functions. The collection is intended to support property prediction, cross-level learning, conformational modeling, and research connecting quantum features with annotated biological data.
Background & Summary
- Curated, comprehensive quantum-chemical datasets remain needed, while computational costs limit the chemical-space scope and modeling applicability of existing collections.The paper also identifies scarce datasets for delta-learning and for three-dimensional conformational data beyond classical force-field levels.Source: Background & Summary · p. 1
- QMugs adds quantum mechanical properties and wave functions for more than 665k ChEMBL-derived molecules, with about 2.0M conformers and over 7 terabytes of data.It provides GFN2-xTB and DFT-level information, including DFT density and orbital matrices.Source: Background & Summary · pp. 1–2
Methods
- The workflow extracts molecules from ChEMBL version 27 and uses RDKit, GFN2-xTB, and Psi4 for conformer generation and quantum calculations.Figure 3 gives an overview of the processing pipeline.Source: Methods · p. 2
Data extraction and SMILES processing
- 685,917 ChEMBL molecules with qualifying assay information were filtered, neutralized, and stripped of salts and solvents; structures outside 3–100 heavy atoms and non-neutral or radical species were removed.The dataset includes ten atom types, from hydrogen through iodine.Source: Data extraction and SMILES processing · p. 2
Conformer generation and optimization
- Conformer generation uses ETKDG, force-field minimization, metadynamics, and RMSD-based clustering to select three conformations per molecule for GFN2-xTB optimization.The optimized structures also have thermodynamic quantities extracted; structures with imaginary vibrational frequencies undergo further optimization, up to 100 attempts.Source: Conformer generation and optimization · pp. 2–3
Quantum mechanical calculations
- DFT single-point calculations on optimized geometries provide molecular and atomic properties alongside density and orbital matrices; 52 structures were discarded after unsuccessful or resource-intensive calculations.Calculations use ωB97X-D/def2-SVP in Psi4.Source: Quantum mechanical calculations · p. 3
Data Records
- Computed molecular structures, properties, and wave functions are available through the ETH Library Collection service.The paper identifies this repository as Data Citation 1.Source: Data Records · p. 3
Format specification
QMugs distributes molecular properties and wave-function files alongside structural data, and evaluates geometry quality and agreement between GFN2-xTB and DFT. The analyses report geometry changes across the pipeline and cross-level correlations for single-point properties.
Format specification
- The SDFs contain single-point properties at both theory levels; separate files provide vibrational spectra and DFT density and orbital matrices in downloadable NumPy tarballs.The wave-function files are split into 100 compressed tarballs of about 50 GB each, with molecules assigned at random to support subset downloads.Source: Format specification · p. 3
Optimized geometry sanity checks
- Four geometry checks screened bond-length deviations, molecular-graph isomorphism, triple-bond linearity, and aromatic-ring planarity; 10,986 conformations were discarded.The cutoffs included 0.2 Å for bond-length deviation, 10° for triple-bond deviation, and 15° for aromatic-carbon dihedral angles.Source: Optimized geometry sanity checks · pp. 3–4
Further geometrical assessment
- Mean pairwise RMSD was 2.40 ± 0.52 Å after MTD sampling, 2.67 ± 0.74 Å after k-means selection, and 2.48 ± 0.86 Å after optimization.For 0.09% of molecules, optimized conformers had mean pairwise RMSD < 0.01 Å.Source: Further geometrical assessment · p. 4
- The geometry shift during GFN2-xTB optimization tended to increase with pre-optimization conformer diversity, while its variance also increased.The authors describe this relationship as heteroscedastic.Source: Further geometrical assessment · p. 4
- For 500 randomly selected structures, independently optimized GFN2-xTB and ωB97X-D/def2-SVP geometries differed by 0.47 ± 0.63 Å RMSD.Structures that did not converge within 100 iterations or the available computational resources were discarded; large RMSDs may indicate distinct local minima.Source: Further geometrical assessment · p. 4
- For 2,067 molecules, individual conformers had SMILES strings describing two different (E)/(Z) isomers; 17,176 molecules shared a SMILES string with another ChEMBL-ID.The possible isomerization examples include α-β-unsaturated ketones, α-β-unsaturated nitriles, imines, and azo compounds.Source: Further geometrical assessment · p. 5
Validation of single-point properties
- PCCs were 0.998 for EForm, 0.969 for µ, 0.999 for A, and 0.769/0.924/0.830 for HOMO/LUMO/gap across GFN2-xTB and DFT.For partial charges, 7 of 10 atom types had PCCs > 0.8; for bond orders, 10 of 15 had PCCs > 0.9, with carbon–fluorine at 0.153.Source: Validation of single-point properties · p. 5
Usage Notes; Code availability
- SDFs can be processed with cheminformatics software, wave-function files with Psi4 or NumPy, and analyses used Python 3.7.7, RDKit 2020.03.3, GFN2-xTB 6.3.1, and Psi4 1.3.2.The archive readme provides instructions for browser access and programmatic downloads.Source: Usage Notes · p. 5, Code availability · p. 5
Figures & Tables
The figures and tables describe QMugs' molecular and property distributions, its data-generation pipeline, and comparisons across theory levels. They also summarize dataset statistics and the properties stored in structure and wave-function files.
Figures & Tables
- Figure 1A plots normalized principal moments; panel B compares unique-InChI overlap between QMugs and QM918, PubChemQC, and ANI-1.The caption notes that repeated InChI strings mean the overlap counts do not match Table 1 totals.Source: Figure 1 · p. 8
- Figure 2 shows distributions of properties for molecules in the QMugs dataset.Source: Figure 2 · p. 9
- Table 1 places QMugs descriptive statistics alongside other DFT-level molecular datasets and the information each provides.Its heavy-atom averages are weighted by conformer count.Source: Table 1 · p. 9
- Figure 3 traces molecules from ChEMBL extraction and standardization through RDKit starting conformers, three MTD conformations, geometry checks, and DFT property calculations.DFT properties use ωB97X-D/def2-SVP and Psi4.Source: Figure 3 · p. 10
- Figure 4 compares conformer RMSD across pipeline stages, geometry shifts against initial diversity, and structures optimized with GFN2-xTB versus ωB97X-D/def2-SVP.The caption reports higher diversity after k-means sampling and reduced diversity after geometry optimization.Source: Figure 4 · p. 11
- Figure 5 compares molecular properties computed with GFN2-xTB and ωB97X-D/def2-SVP, including formation energy, dipole moment, and rotational constant A.Twenty-two small-molecule conformations with very large rotational constants are omitted; including them gives A RMSE 845.834 cm−1 and PCC 0.091.Source: Figure 5 · p. 12
- Figure 6 compares partial-charge correlations by atom type, while Figure 7 compares Wiberg bond orders for the 15 most frequent bond types.For bond types occurring more than 1M times, Figure 7 plots a randomly selected sample of 1M bonds.Source: Figure 6 · p. 12, Figure 7 · p. 13
- Table 2 lists calculated properties in QMugs SDFs and flags those enabling delta-learning; Table 3 lists wave-function-file molecular properties and gives Mayer and Wiberg-Löwdin bond orders as a superset of SDF bond orders.Source: Table 2 · p. 14, Table 3 · p. 14
1 Overlap with other datasets
The section compares QMugs compounds with those in QM9, PubChemQC, and ANI-1 using InChI strings to assess dataset overlap.
1 Overlap with other datasets
- Compound overlap was assessed between QMugs and QM9, PubChemQC, and ANI-1.The comparison used the compounds’ InChI strings.Source: §1 · p. 15
- Molecules from each dataset’s original format were converted to InChIs with Open Babel 3.1.1; failed conversions were skipped.The original formats included SDF, XYZ, MOL, and SMILES.Source: §1 · p. 15
- Different washing procedures mean one compound identifier can correspond to different InChI strings across datasets.Protonated and unprotonated forms are one stated example.Source: §1 · p. 15
2 ChEMBL SQL query
The ChEMBL query selected single-protein targets and compounds with qualifying activity annotations, then converted and filtered the activity values.
2 ChEMBL SQL query
- The extraction queried ChEMBL in two stages: first targets, then compounds with annotated activity toward those targets.A locally downloaded MySQL instance of ChEMBL 27 was queried.Source: §2 · p. 15
2.1 Extraction of targets
- Targets required qualifying activity types and annotations, with assay relationship, confidence, mutant, duplicate, validity, and missing-data filters applied.Only targets retaining annotations for at least 10 unique compounds were kept.Source: §2.1 · p. 15
2.2 Extraction of compounds
- Compound annotations missing activity values or units were discarded, and remaining values were converted to pX and restricted to 3–12.Inactivity annotations were assigned pX = 3; values below -3 before correction were converted to their absolute value.Source: §2.2 · p. 16
3 SMILES filtering
The SMILES filtering stage removed molecules containing specified substructures that conflicted with the force field or were considered less relevant to drug-like molecules.
3 SMILES filtering
- Molecules containing a specified set of SMARTS patterns were removed using RDKit.The listed patterns included boron and a range of elements.Source: §3 · p. 16
- The excluded substructures included groups incompatible with MMFF94s, such as boron, and halogens bonded to two neighbors.Elements considered less relevant for drug-like molecules were also excluded.Source: §3 · p. 16
4 Computational resources
Computationally demanding or unsuccessful calculations led to a small number of structure exclusions.
4 Computational resources
- DFT calculations exceeding 4 CPU cores for 24 hours, 40 GB memory, or 100 GB scratch space were excluded; this applied to 9 single-point calculations.A further 19 GFN2-xTB and 24 DFT structures were discarded because calculations could not be completed successfully.Source: §4 · p. 16
5 Optimized geometry sanity checks
Four consecutive geometry checks removed structures with unrealistic optimized conformations; some conformations failed multiple checks, so failure counts do not sum to the total removed. The checks assessed bond lengths, molecular graph isomorphism, non-ring triple-bond linearity, and aromatic-ring planarity.
5.1 Deviation of bond lengths from experimental reference values
- 6,131 conformations failed the bond-length test, which discarded molecules whose largest deviation from reference bond lengths exceeded 0.2 Å.Reference values were average bond lengths between atom types from a computational chemistry database; the threshold was selected by visual and manual inspection.Source: §5.1 · pp. 16–17
- QMugs conformations had a median largest bond-length deviation of 0.0687 ± 0.0317 Å, compared with 0.0580 ± 0.0419 Å for PubChemQC.The QMugs sample was larger on average, and the authors note that greater steric clashes in larger molecules may contribute to the slightly higher deviations.Source: Figure S1 · p. 17
5.2 Molecular graph isomorphism
- 1,568 conformations failed the molecular graph isomorphism test after bond information was removed from the structure files.The comparison labelled graph nodes by atom type but omitted bond types because of false negatives associated with resonance structures and other cases.Source: §5.2 · pp. 17–18
5.3 Deviation of triple bonds from linear geometry
- 1,147 conformations failed the non-ring triple-bond check, which discarded structures with angular deviation Δγ > 10° from linear geometry.The check excluded ring triple bonds because ring strain can produce substantial departures from linearity.Source: §5.3 · p. 18, Figure S2 · p. 18
5.4 Deviation of aromatic rings from planar geometry
- 2,769 conformations failed the aromatic-ring check, which discarded structures with an aromatic-carbon dihedral deviation of 15° or more.The QMugs conformations showed a median deviation of 2.99 ± 2.20°, versus 1.70 ± 1.85° for PubChemQC.Source: §5.4 · pp. 18–19, Figure S3 · p. 19
6 Independent terms of the Schrödinger equation
The Born Oppenheimer approximation represents molecular energy in the electronic Schrödinger equation as four independent terms. Figure S4 visualizes the distributions of those terms for QMugs molecules at the ωB97X-D/def2-SVP level.
6 Independent terms of the Schrödinger equation
- Molecular energy is expressed as the sum of nuclear repulsion, exchange-correlation, one-electron, and two-electron terms.The one-electron term is also called kinetic electron energy, and the two-electron term electron repulsion energy.Source: §6 · p. 20
- Figure S4 plots the nuclear repulsion and exchange-correlation energy terms for QMugs molecules in E_H.These are panels A and B, respectively.Source: Figure S4 · p. 20
- Figure S4 also plots the one-electron and two-electron energy terms for QMugs molecules in E_H.These are panels C and D, respectively.Source: Figure S4 · p. 20
7 Thermodynamic properties
QMugs thermodynamic properties were calculated at the GFN2-xTB level of theory. The reported properties include molecular free energy, enthalpy, Fermi level, heat capacity, and temperature-dependent entropy.
7 Thermodynamic properties
- The thermodynamic properties in QMugs were calculated using GFN2-xTB.Source: §7 · p. 21
- The reported properties include total Gibbs free energy G, total enthalpy H, Fermi level E_Fermi, and total heat capacity C_Temp.Figure S5 shows distributions for G, H, E_Fermi, and C_Temp; the caption gives eV for the first three.Source: §7 · p. 21, Figure S5 · p. 21
- Temperature-dependent entropy S_Tot and the other reported thermodynamic properties include rotational, translational, and vibrational contributions that can be found in the SDF.The corresponding distributions are shown in Figure S5.Source: §7 · p. 21
8 Additional figures and tables
The additional tables compare QMugs atomic partial charges and Wiberg bond orders across GFN2-xTB and ωB97X-D/def2-SVP. They report comparison metrics including RMSE and Pearson’s correlation coefficient.
8 Additional figures and tables
- Table 1 compares atom-type-specific atomic partial charges between GFN2-xTB and ωB97X-D/def2-SVP for QMugs.The table identifies RMSE and Pearson’s correlation coefficient as comparison metrics.Source: Table 1 · p. 21
- Both tables report RMSE and Pearson’s correlation coefficient for comparisons between the two levels of theory.Source: Table 1 · p. 21, Table 2 · p. 22
- Table 2 compares Wiberg bond orders between GFN2-xTB and ωB97X-D/def2-SVP for the 15 most frequent pair-wise atomic covalent bonds in QMugs.The table identifies RMSE and Pearson’s correlation coefficient as comparison metrics.Source: Table 2 · p. 22