Source-linked AI summary
GEOM: Energy-annotated molecular conformations for property prediction and molecular generation
Simon Axelrod, Rafael Gomez-Bombarelli
TL;DR
Existing 2D graphs and single 3D structures do not capture molecular flexibility or the conformer ensembles governing finite-temperature properties. GEOM addresses this gap with large-scale, high-quality conformer generation connected to experimental data, while validating ensemble quality and demonstrating improved ensemble-property prediction.
Problem
2D graphs and single 3D structures fail to capture molecular flexibility, although molecular properties depend on conformers accessible at finite temperature.
Method
GEOM combines extensive CREST sampling with the semi-empirical GFN2-xTB method to generate high-quality conformers for molecules with experimental data and QM9 molecules.
Results
GEOM provides high-quality conformers for 317,928 mid-sized organic molecules with experimental data and 133,258 QM9 molecules, and 3D ensemble features reduce prediction error versus ChemProp for G and ln(unique conformers).
Takeaways & Limitations
GEOM connects large-scale conformer ensembles with experimental measurements and supports models that use molecular flexibility for property prediction and molecular generation.
Takeaways & Limitations
Full DFT optimization and accurate free-energy calculation for every ensemble is computationally daunting because ensembles may contain hundreds of conformers and require Hessian calculations.
Abstract
from arXiv · showhide
Machine learning (ML) outperforms traditional approaches in many molecular design tasks. ML models usually predict molecular properties from a 2D chemical graph or a single 3D structure, but neither of these representations accounts for the ensemble of 3D conformers that are accessible to a molecule. Property prediction could be improved by using conformer ensembles as input, but there is no large-scale dataset that contains graphs annotated with accurate conformers and experimental data. Here we use advanced sampling and semi-empirical density functional theory (DFT) to generate 37 million molecular conformations for over 450,000 molecules. The Geometric Ensemble Of Molecules (GEOM) dataset contains conformers for 133,000 species from QM9, and 317,000 species with experimental data related to biophysics, physiology, and physical chemistry. Ensembles of 1,511 species with BACE-1 inhibition data are also labeled with high-quality DFT free energies in an implicit water solvent, and 534 ensembles are further optimized with DFT. GEOM will assist in the development of models that predict properties from conformer ensembles, and generative models that sample 3D conformations.
Background & Summary
GEOM addresses the lack of large-scale datasets that combine molecular conformer ensembles with experimental properties, providing high-quality conformers across diverse molecular datasets. It also supports conformer-ensemble property prediction and generative modeling of 3D conformations.
- Dataset contribution: GEOM provides high-quality conformers for 317,928 mid-sized organic molecules with experimental data and 133,258 QM9 molecules.The experimental data cover physical chemistry, biophysics, and physiology, including biological assay results.
- Dataset contribution: GEOM combines molecular graphs, conformer ensembles, and experimental data to address a gap left by datasets containing only 2D graphs or single 3D structures.Existing datasets may lack full ensembles, scale, conformer quality, or experimental connections.
- Applications: GEOM supports training models that predict experimental properties from conformer ensembles, including biological, physicochemical, and physiological properties.These models could not be trained on datasets containing only 2D graphs or single 3D structures.
- Applications: GEOM supports generative models that predict molecular conformers from input molecular graphs and benchmarks their recall and diversity.Its size and simulation accuracy make it suitable for training and benchmarking conformer-generation methods.
- Construction and scope: 98% of the original molecules in nearly all MoleculeNet datasets have vacuum conformer ensembles, but missing molecules may bias comparisons because many are extremely flexible.Some conformer calculations ran for days without finishing, while others failed during xTB optimization.
CREST
CREST combines enhanced sampling with semi-empirical energies to generate broad conformer ensembles, then ranks and refines them for higher-quality statistical weights. GEOM uses these ensembles, with selected DFT calculations, to support property prediction and conformer-generation models.
- Sampling: GFN2-xTB provides a practical speed–accuracy compromise for generating initial conformer ensembles, while full DFT is too costly for exhaustive ensemble optimization.CREST can perform millions of energy and force calculations, making xTB practical where full DFT is prohibitive.
- Energy ranking: CREST assigns approximate Boltzmann statistical weights from conformer energies and degeneracies, but omits conformation-dependent free-energy contributions from translation, rotation, and vibration.The approximation can mis-rank conformers because statistical weights depend exponentially on energy differences.
- Sampling: CREST uses metadynamics-like simulations with bias potentials to drive molecules across torsional barriers and explore new conformations.The bias is based on RMSD collective variables, and new reference structures are added during trajectories to encourage continued exploration.
- Ensemble processing: Conformers, rotamers, and duplicates are distinguished using energy differences, RMSDs, and rotational-constant differences, followed by additional molecular-dynamics sampling of low-energy barriers.The lowest-energy structures undergo simulations at 400 K and 500 K to identify remaining rotamers.
- Validation: CREST ensembles show broad conformational coverage: all 15 lowest-energy experimental citronellal conformers were recovered.Ensemble-based NMR calculations matched experiment better than calculations using only one conformer, with similar results reported for several other molecular systems.
- GEOM application: GEOM applies this workflow to large molecular collections, adding DFT energies for 1.3 million CREST conformers across 1,511 BACE species and fully optimizing 534 species.The resulting ensembles support conformer-ensemble property models, single-conformer models, and transfer learning for conformer generation.
Conformer generation
GEOM conformers were generated from standardized molecular inputs using RDKit, xTB, CREST, and selected DFT optimizations, followed by graph re-identification and additional BACE calculations.
- SMILES strings were canonicalized with RDKit, while salts and clusters were processed to identify the molecular component and standardize representations.
- Initial geometries were generated and optimized with RDKit, MMFF, and xTB before selecting the lowest-energy structure as the CREST seed.
- CREST generated conformer ensembles from one xTB-optimized structure per species, with computational costs varying substantially for flexible molecules.
- Graph re-identification was required because stereochemistry was implicit in generated geometries and CREST could produce dissociation or tautomerization.
- xyz2mol converted conformers into RDKit molecular objects for graph-feature assignment, although resonance structures could yield inconsistent graphs.
- 534 BACE molecules were further optimized with CENSO, while all BACE CREST conformers received single-point DFT and Hessian calculations in water.
Conformational property prediction
GEOM supports prediction of ensemble-defined conformational properties from molecular representations. Models were trained to predict free energy, average energy, and conformer counts, with graph-based models performing best overall.
- GEOM enables property predictors and generative models to use conformer ensembles, including tasks that test graph-based inference of flexibility-mediated properties.
- The predicted quantities were conformational free energy, average conformational energy, and the number of unique CREST conformers.
- Models included KRR, random forest, FFNN, ChemProp, and SchNetFeatures, using fingerprints, graphs, or a highest-probability 3D conformer.
- 100,000 AICures drug species were randomly split 60-20-20 into training, validation, and test sets for each model.
- ChemProp and SchNetFeatures were strongest overall; relative to ChemProp, SchNetFeatures reduced error 10% for G and 5% for ln(unique conformers), with nearly equal error for ⟨E⟩.
- GEOM-trained models can predict expensive simulation results, while generative models can produce conformers themselves.
Data Records
GEOM is distributed through multiple file formats and organizes species-level summaries alongside conformer-level structural information.
- The dataset is available online with tutorials for loading and analyzing the data.
- MessagePack files store species dictionaries keyed by SMILES strings, including experimental and conformer summary information.
- Python pickle files are organized by dataset and calculation type, with one file per species containing summary and conformer information.
Technical Validation
Technical validation checked parsing accuracy, graph recovery, and agreement between CREST and higher-level DFT calculations. The results support broad conformer validity while identifying energy-ranking limitations.
- Manual inspection confirmed that sampled conformer data from 20 species matched the original CREST output files.
- Graph re-attribution succeeded for 88.4% of QM9 molecules and 94.7% of drug molecules, with manual inspection suggesting 98.4% original-graph recovery for drugs.
- CREST xTB energies had 1.96 kcal/mol MAE and moderate within-species ranking correlation with r2scan-3c DFT, averaging ρ = 0.39 ± 0.35.
- Single-point versus optimized DFT energies showed average Spearman correlation 0.69 and MAE 0.54 kcal/mol, 3.6 times lower than the xTB error.
- CENSO energy and free-energy rankings correlated at ρ = 0.85 ± 0.18 with 0.33 kcal/mol MAE, indicating energies can order conformers by statistical weight.
Usage Notes
The authors provide tutorials and recommend RDKit pickle files for convenient dataset loading and analysis, while directing 3D-model training users to the tutorial and README.
- RDKit pickle files are recommended because RDKit mol objects are easy to handle and their properties can be readily analyzed.
- MessagePack files are secure and language-independent but more difficult to analyze because they represent graphs through features rather than objects with built-in methods.
- The authors suggest following the tutorial and README to train 3D-based models.
Code availability
Dataset-loading tutorials and 3D-model training code are publicly available without restriction, alongside the computational tools used to generate and refine the ensembles.
- Dataset-loading tutorials and code for training 3D-based neural network models are publicly available without restriction.
- CREST and xTB are freely available online, and specific software versions were used for ensemble generation, refinement, and single-point calculations.
- CREST 2.9 with xTB 6.2.3 generated initial conformer ensembles, while CENSO 1.1.2 with Orca 5.0.1104 and xTB 6.4.1 refined them.
- Orca 5.0.2 was used for all single-point calculations.
Author contributions statement
R.G.B. conceived the project, S.A. performed the calculations, and both authors wrote and revised the manuscript.
- R.G.B. conceived the project, while S.A. performed the calculations.
- Both authors wrote and revised the manuscript.
Competing interests
The authors declare that they have no competing interests.
- The authors declare no competing interests.
Figures & Tables
The figures and tables present GEOM’s molecular representations, dataset statistics, experimental coverage, conformer-related prediction results, and energy-ranking comparisons.
- Molecular representations: Figure 1 combines a SMILES string, a stereochemical formula with edge features, and an overlay of molecular conformers.In the conformer overlay, higher transparency indicates lower statistical weight.
- Experimental coverage: Table 3 lists MoleculeNet species represented in GEOM and their recovery percentages, including 1,511 BACE species with CREST conformers in water.The BACE recovery is 99.9%; duplicate compounds after preprocessing explain differences from original MoleculeNet counts.
- Prediction results: Table 4 compares mean absolute errors for three conformer-related properties predicted from the AICures drug dataset.ChemProp and SchNetFeatures are strongest overall, followed by FFNN, KRR, and random forest.
- Energy and geometry comparisons: Figures 4, 5, and 6 compare CREST or GFN2-xTB rankings and geometries with higher-level DFT or CENSO calculations.Figure 4 evaluates xTB energy ranking against r2scan-3c energies, while Figures 5 and 6 assess CENSO rankings and geometry deviations.