A conformational benchmark for optical property prediction with solvent-aware graph neural networks

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Accurate prediction of molecular optical spectra is crucial for developing advanced materials like OLED emitters, solar-cell dyes, and fluorescent probes.
• Traditional computational methods (e.g., time-dependent density-functional theory) are often too slow and inaccurate for practical applications.
• Existing Graph Neural Network (GNN) models, while faster, typically use 2D graphs and neglect essential 3D geometric information influencing excited-state properties.

Purpose of the Study:

• To develop a high-quality dataset (nablaColors-3D) for training and evaluating 3D GNNs on molecular optical property prediction.
• To establish a robust benchmark for 3D GNNs, assessing the impact of geometry optimization fidelity on prediction accuracy.
• To propose and validate a novel solvent-aware 3D GNN architecture for enhanced optical spectra prediction.

Main Methods:

• Curated the nablaColors-3D dataset, comprising 26,369 chromophore-solvent pairs with multi-level quantum mechanically optimized geometries.
• Developed a scaffold-split benchmark to rigorously evaluate 3D GNN performance, isolating the effects of molecular structure.
• Proposed a solvent-aware modification for SE(3)-invariant GNN architectures, leveraging pretrained models like UniMol+.

Main Results:

• The nablaColors-3D dataset and benchmark provide a standardized platform for advancing 3D GNNs in computational chemistry.
• Systematic analysis revealed the significant impact of geometry optimization quality on the accuracy of predicted optical spectra.
• The best-performing solvent-aware 3D GNN model achieved a Mean Absolute Error (MAE) of 15.97 nm on a held-out test set, surpassing previous state-of-the-art results by over 30%.

Conclusions:

• 3D Graph Neural Networks, when trained on comprehensive datasets like nablaColors-3D and incorporating solvent effects, offer a powerful and accurate alternative to traditional methods for predicting molecular optical spectra.
• The developed benchmark and solvent-aware models represent a significant advancement in the application of machine learning for materials discovery in optics and photonics.
• This work paves the way for faster and more accurate design of molecules for OLEDs, solar cells, and sensing applications.