Reproduce “GEOM: Energy-annotated molecular conformations for property prediction and molecular generation” on OpenResearch

About this paper

Simon Axelrod, Rafael Gomez-Bombarelli

GEOM is a comprehensive dataset featuring 37 million energy-annotated molecular conformations for over 450,000 distinct molecules, including QM9 and drug-like compounds linked to experimental data. The dataset was constructed using multi-level quantum chemical methods, providing accurate 3D conformer ensembles and associated energies to support machine learning applications in property prediction and generative chemistry.

Key insights

• The CREST (GFN2-xTB) method, balancing speed and accuracy, efficiently explores conformational space, generating initial geometries that are close to high-level DFT minima (mean RMSD of 0.36 Å).
• While GFN2-xTB relative energies have a moderate mean absolute error of 1.96 kcal/mol compared to single-point DFT, using single-point DFT on CREST geometries significantly improves accuracy, achieving an average MAE of 0.54 kcal/mol compared to fully optimized DFT energies.
• Machine learning models, particularly those leveraging 3D geometric information like SchNetFeatures, demonstrate improved accuracy in predicting conformer-derived properties from GEOM, showing that incorporating conformational data enhances prediction capabilities for flexibility-sensitive properties.

Summary by alphaXiv.