GEOM: Energy-annotated molecular conformations for property prediction and molecular generation

Motivation

• Because molecular properties depend on conformers accessible at finite temperature, 2D graphs and single 3D structures omit molecular flexibility and motivate ensemble-based prediction.The molecules are continually moving on a potential energy surface, while their accessible conformers are not explicitly represented by 2D or single-structure inputs.Source: Background & Summary · p. 1

Dataset construction

• 317,928 experimentally characterized molecules and 133,258 QM9 molecules receive high-quality conformer ensembles; 1,511 BACE ensembles carry DFT and xTB energy annotations, and 534 received DFT geometry optimization.The BACE ensembles include single-point DFT energies and semi-empirical quasi-harmonic free energies.Source: Background & Summary · p. 1

Property prediction

• GEOM enables property-prediction benchmarks using conformers and experimental data, combining scale, ensemble size, conformer quality, and experimental linkage absent from prior datasets.Existing datasets either provide single structures at scale, ensembles for only hundreds of molecules, or force-field-quality conformers for tens of thousands of experimentally labeled molecules.Source: Background & Summary · pp. 1–2

Conformer generation

• GEOM supports training and benchmarking conformer-generation models, but CREST conformer coverage is stronger than its statistical weights, so probability benchmarks should use GEOM’s DFT weights.Models trained on GEOM may reproduce its conformer-generation accuracy on unseen molecules at a fraction of the computational cost; CREST ensembles have high coverage of thermally accessible conformers.Source: Background & Summary · p. 2

Experimental data

• AICures data include SARS-CoV-2 and 3CL-protease inhibition and bacterial growth assays, while MoleculeNet adds three physical-chemistry tasks, 659 physiology tasks, and BACE-1 binding affinities.The 3CL protease’s 96% sequence similarity to its SARS-CoV counterpart makes the data relevant to transfer-learning benchmarks.Source: Background & Summary · p. 2

Coverage and limitations

• Over 98% of molecules in each dataset except SIDER received vacuum ensembles, and water-model ensembles covered 99.9% of BACE; missing compounds may be unusually flexible or have unusual topologies, biasing comparisons.Some flexible calculations ran for several days without finishing, while others failed during initial xTB optimization.Source: Background & Summary · p. 2

CREST

• CREST addresses sparse sampling and inaccurate force-field energies by pairing metadynamics exploration with semi-empirical tight-binding DFT, which is more accurate than force fields and cheaper than full DFT.Its energy-based grouping also identifies rotamers that differ only by atom re-indexing and assigns conformer probabilities using degeneracy and energy.Source: CREST · pp. 2–3
• CREST runs NVT metadynamics for 5–200 ps, adds trajectory structures as RMSD-based bias references every 1.0 ps, and uses 12 runs with different bias settings to explore conformational space.The bias promotes torsional barrier crossings rather than bond-breaking, while the bare energy limits exploration of ultra-high-energy regions and thermally inaccessible structures.Source: CREST · p. 3
• After GFN2-xTB optimization, CREST classifies conformers, rotamers, and duplicates using energy, RMSD, and rotational-constant thresholds, then iterates sampling and refinement before retaining structures within a 6.0 kcal/mol window.The workflow includes up to five initial restarts, 400 K and 500 K MD, genetic Z-matrix crossing, tight optimization, and further restarts when lower-energy conformers are found.Source: CREST · p. 3
• CREST recovered all 15 lowest-energy experimental gas-phase citronellal conformers, and the ensemble-based computed 1H-NMR spectrum matched experiment far better than a single-conformer spectrum.The cited work also reported similarly good results for macrocycles, a protonated peptide, metal-organic systems, and the 1-Naphthol dimer.Source: CREST · pp. 3–4
• CREST’s Boltzmann-like probabilities omit translation, rotation, and vibrational free energies, while typical xTB errors of 2 kcal/mol exceed room-temperature kBT of 0.59 kcal/mol and limit accurate ranking.Full DFT optimization and Hessian-based corrections are resource-intensive for ensembles containing hundreds of conformers, although CREST’s 6.0 kcal/mol window is intended to retain accessible structures.Source: CREST · pp. 3–4
• CENSO progressively tightens DFT optimization and free-energy cutoffs, uses r2scan-3c for final optimization, and computes entropic and zero-point corrections with biased xTB Hessians.Unlike CREST, CENSO does not include rotamer degeneracy; its free energies include gas-phase energy, solvation free energy, and translation, rotation, and vibration contributions.Source: CREST · p. 4
• CENSO-generated conformers and statistical weights qualitatively reproduce solution-phase optical rotation and yield accurate octanol-water partition coefficients and pKa values.The paper describes the resulting conformer ensembles and statistical weights as reliable.Source: CREST · p. 4
• CENSO ensembles were generated for 534 species, while single-point DFT energies and xTB vibrational frequencies cover 1,511 of 1,513 BACE species (99.9%).The CENSO calculations use implicit water solvent; the DFT-annotated ensembles can support binding-affinity modeling and transfer learning.Source: CREST · pp. 4–5

SMILES pre-processing

• Drug-like inputs were canonicalized and de-salted to standardize species and ionization, while MoleculeNet clusters were reduced to their heaviest component without protonation.For non acid-base clusters, the compound of interest was the heaviest component; acid/base SMILES were converted to protonated molecule and counterion. Original SMILES are retained as uncleaned_smiles.Source: SMILES pre-processing · p. 5

Initial structure generation

• Drug molecules were seeded for CREST from the lowest-energy xTB-optimized RDKit conformer, while QM9 geometries were re-optimized with xTB before sampling.RDKit generated 50 conformers (up to 500 if none succeeded), MMFF-optimized them, removed duplicates below 0.1 Å RMSD, and selected the ten lowest-energy structures for xTB optimization; MMFF optimization was skipped for SMILES indicating cis/trans stereochemistry.Source: Initial structure generation · p. 5

CREST simulation

• CREST used one xTB-optimized seed structure per species with default arguments except geometry charge, and its computational cost varied by dataset and hardware.Average wall time was 2.8 hours on 32 cores for AICures drug jobs on KNL nodes, 0.63 hours on 13 cores on Cascade Lake/Sky Lake, 0.04 hours on 13 cores for QM9, and 6.3 hours on 18.1 cores for MoleculeNet jobs; the MoleculeNet median was 1.4 hours on 12 cores.Source: CREST simulation · p. 5

Graph re-identification

• Conformer graphs were re-identified because stereochemistry may be unspecified in input SMILES and CREST sampling may produce dissociation or tautomerization.xyz2mol generated RDKit molecules for graph features, but could assign resonance structures that made conformers of one species appear to have different graphs, distinct from reactivity-driven graph changes.Source: Graph re-identification · pp. 5–6

CENSO simulation

• 534 BACE molecules (35%) were refined with CENSO using DFT in implicit water, after CREST sampling with an ALPB water model.Conformer and rotamer duplicates were removed throughout optimization; the average CENSO job took 1 day and 4 hours on 54 cores, totaling 781,000 CPU hours.Source: CENSO simulation · p. 6

Single point calculations

• Single-point DFT calculations were performed without further optimization on all 1.3 million CREST conformers in the BACE dataset.The calculations used r2scan-3c/mTZVPP with C-PCM water and took 1.1 million CPU hours in total.Source: Single point calculations · p. 6

Hessian calculations

• Hessian calculations used xTB with ALPB water for all 1.3 million CREST conformers in the BACE dataset.The calculations took 63,000 CPU hours in total.Source: Hessian calculations · p. 6

Applications

• GEOM supports conformer-based property prediction and generative modeling; the benchmark illustrates predicting costly simulation results, while generating CREST structures from scratch is too costly for virtual screening.The authors note that GEOM-trained generative models could also produce conformers themselves.Source: Conformational property prediction · pp. 6–7

Tasks

• The benchmark predicts conformational Gibbs free energy, average conformational energy, and CREST’s number of unique conformers within its default maximum energy window.These ensemble-defined properties assess models’ ability to infer properties mediated by conformational flexibility from a molecular graph.Source: Conformational property prediction · p. 6

Models

• The study compares fingerprint-based random forest, KRR, and FFNN models with graph-based ChemProp and SchNetFeatures, which uses graph features and 3D geometry.SchNetFeatures was trained on each species’ highest-probability conformer.Source: Conformational property prediction · pp. 6–7

Training setup

• Models were trained on a random sample of 100,000 GEOM AICures drug-subset species using a shared 60-20-20 train-validation-test split.Hyperparameters were optimized separately for each model type and task.Source: Conformational property prediction · p. 7

Results

• ChemProp and SchNetFeatures are strongest overall; versus ChemProp, SchNetFeatures’ error is 10% lower for G, nearly equal for ⟨E⟩, and 5% lower for ln(unique conformers).The authors attribute the limited added value of one 3D geometry to ensemble properties being mainly determined by molecular flexibility, which is encoded by graph features such as rotatable bonds.Source: Conformational property prediction · p. 7

Data Records

• The dataset is available online, with tutorials for loading and analyzing the data (118, 97).Source: Data Records · p. 7
• MessagePack files for AICures and QM9 separate crude species and conformer information from featurized conformer molecular graphs.Each file maps SMILES strings to sub-dictionaries; crude files include species-level data and conformer geometry, energy, and degeneracy.Source: Data Records · p. 7
• Python pickle files store each species’ summary and conformers as RDKit mol objects, while a lightweight summary JSON supports filtering species before loading pickles.The folders cover QM9, AICures, MoleculeNet, and BACE calculations in water and with CENSO.Source: Data Records · p. 7

• Manual checks confirmed that one randomly sampled conformer from each of 20 species matched its data in the CREST output files.Source: Technical Validation · p. 7
• Graph re-identification recovered 88.4% of QM9 and 94.7% of drug graphs; accounting for resonance-form failures, the original graph was likely recovered for 98.4% of drugs.All failed QM9 graphs underwent reactions, while manual inspection suggested that 70% of 53 AICures drug cases failed because of xyz2mol resonance-form handling.Source: Technical Validation · p. 7
• For BACE conformers, xTB had a 1.96 kcal/mol MAE; within-species energy-ranking ρ averaged 0.39 ± 0.35, indicating moderate correlation with single-point DFT.Across all geometries from all species, ρ was 0.47. The correlation was significantly better than for classical force fields; MMFF94’s median ρ versus single-point DFT was between −0.1 and −0.45 for drug-like molecules.Source: Technical Validation · p. 8
• SP and CENSO conformer energies had a 0.54 kcal/mol MAE and average Spearman ρ of 0.69 ± 0.27; mean geometry RMSD was 0.36 Å.The energy MAE was 3.6 times lower than the xTB error. Comparisons included only conformers with non-negligible weight after optimization (ΔG ≤ 2.5 kcal/mol), because CENSO discards high-energy conformers.Source: Technical Validation · p. 8
• For lowest-energy CENSO geometries, CREST–CENSO rank ρ was 0.43 ± 0.41; within CENSO, energy–free-energy rank ρ was 0.85 ± 0.18 and MAE was 0.33 kcal/mol.Only the lowest-energy CENSO geometries were included in the CREST–CENSO comparison, unlike the comparison of all CREST conformers with single-point energies. The results indicate that statistical-weight errors in GEOM are dominated by xTB errors, with quasi-harmonic errors comparatively negligible.Source: Technical Validation · p. 8

Usage Notes

• Use RDKit pickle files for data loading and analysis; MessagePack is secure and cross-language accessible but harder to analyze because it represents graph features rather than objects with built-in methods.Researchers are also encouraged to use the data-loading tutorials in [97].Source: Usage Notes · p. 8
• For training 3D-based models, follow the tutorial and README file [117].Source: Usage Notes · p. 8

Code availability

• Tutorials for loading the dataset and code for training 3D-based neural network models are publicly available without restriction.CREST and xTB are also freely available online.Source: Code availability · p. 8
• Initial conformers were generated with CREST 2.9 and xTB 6.2.3, then refined with CENSO 1.1.2, Orca 5.0.1, and xTB 6.4.1.Orca 5.0.2 was used for all single-point calculations.Source: Code availability · p. 8
• After a race-condition bug produced incorrect CENSO energies in Orca 5.0.1, ensembles with failed calculations or energy ranges above 30 kcal/mol were discarded.New Orca 5.0.2 single-point calculations on converged CENSO geometries found and replaced 0.44% incorrect energies.Source: Code availability · p. 8

• R.G.B. conceived the project, S.A. performed the calculations, and both authors wrote and revised the manuscript.Source: Author contributions statement · p. 14

Figures & Tables

• Figure 1 presents latanoprost as a SMILES string, a stereochemical formula with edge features, and an overlay of conformers whose transparency indicates statistical weight.The formula marks in- and out-of-plane bonds with wedges and cis isomerism with a double line.Source: Figure 1 · p. 14
• Tables 1–3 cover molecular descriptor statistics for QM9 and AICures and experimental data for AICures and MoleculeNet species.Table 3 defines species with vacuum CREST conformers and reports recovery as a percentage of original MoleculeNet compounds; it notes 1,511 BACE species (99.9%) also have water CREST conformers.Source: Table 1 · p. 15, Table 2 · p. 15, Table 3 · p. 15
• Figures 2 and 3 present CREST-based statistics for the QM9 and AICures drug datasets, including violin plots in Figure 3.Source: Figure 2 · p. 16, Figure 3 · p. 16
• Table 4 reports prediction mean absolute errors for three conformer-related properties using models trained and tested on the AICures drug dataset.Source: Table 4 · p. 16
• Figures 4–6 compare CREST or CENSO energy rankings with single-point DFT rankings and compare CREST geometries with DFT-optimized geometries.The figures show Spearman rank-correlation distributions; Figure 4 also compares GFN2-xTB and r2scan-3c energies, while Figure 5 reports geometry RMSDs and Figure 6 contrasts energy and free-energy rankings with CENSO.Source: Figure 4 · p. 17, Figure 5 · p. 17, Figure 6 · p. 18