| Dataset Details | |
| Authors | Simon Axelrod, Rafael Gómez-Bombarelli |
| Paper Title | GEOM, energy-annotated molecular conformations for property prediction and molecular generation |
| Institutions | Harvard University, MIT |
| Published In | Nature Scientific Data |
| Category | Computational Chemistry |
| Format | SMILES RDKit mol objects Experimental annotations Computational annotations |
| Size | Conformations: 37,000,000+ Molecules: 450,000+ |
| Date | September 2025 |
| Year | 2022 |
| Links | 📊 Dataset • 🔗 DOI • 📄 Paper |
This paper presents the Geometric Ensemble Of Molecules (GEOM) dataset to address the critical lack of a large-scale dataset linking molecular conformer ensembles to experimental data, which is necessary for training more advanced and accurate machine learning models.
| Type | Count |
|---|---|
| Conformations | 37,000,000+ |
| Molecules | 450,000+ |
| Subset | Count | Description |
|---|---|---|
| Drug-like (AICures) | 304,466 molecules | AICures COVID-19 challenge molecules |
| QM9 | 133,258 molecules | QM9 benchmark dataset |
| MoleculeNet | 16,865 molecules | Molecules from benchmarks for physical chemistry, biophysics, and physiology |
| BACE (High-quality DFT) | 1,511 molecules | A subset of MoleculeNet with DFT energies and experimental data |
| Model | G | E | ln(Conformers) |
|---|---|---|---|
| SchNetFeatures | 🥇 0.203 | 🥈 0.113 | 🥇 0.363 |
| ChemProp | 🥈 0.225 | 🥇 0.110 | 🥈 0.380 |
| FFNN | 0.274 | 0.119 | 0.455 |
| KRR | 0.289 | 0.131 | 0.484 |
| Random Forest | 0.406 | 0.166 | 0.763 |
Semi-empirical conformational sampling with CREST/GFN2-xTB (accuracy level ~2 kcal/mol MAE).
r2scan-3c/mTZVPP energies on CREST geometries (sub-kcal/mol accuracy).
Full r2scan-3c/mTZVPP optimization and free energies (accuracy level ~0.3 kcal/mol vs CCSD(T)).
Conformer generation software. Uses semi-empirical tight-binding DFT to evaluate energies (GFN2-xTB). Runs metadynamics in an NVT ensemble, biasing the potential to explore conformational space. 1.5 million CPU hours consumed in this process.
CENSO applied to 534 species in BACE, yielding high-accuracy ensembles for 35% of the species in the subset. Single-point energies are also completed for the remaining species in BACE. 781,000 CPU hours consumed for CENSO simulations and 1.1 million CPU hours for single-point energy calculations.