Pretrained E(3)-equivariant message-passing neural networks with multi-level representations for organic molecule spectra prediction
Abstract Fast and accurate spectral prediction plays a crucial role in molecular design within fields such as pharmaceutical and materials science. Nevertheless, predicting molecular spectra typically requires quantum chemistry calculations, posing significant challenges for fast predictions and hig...
Saved in:
| Main Authors: | , , , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Nature Portfolio
2025-07-01
|
| Series: | npj Computational Materials |
| Online Access: | https://doi.org/10.1038/s41524-025-01698-z |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract Fast and accurate spectral prediction plays a crucial role in molecular design within fields such as pharmaceutical and materials science. Nevertheless, predicting molecular spectra typically requires quantum chemistry calculations, posing significant challenges for fast predictions and high-throughput screening. In this paper, we propose an equivariant, fast, and robust model, named EnviroDetaNet, which integrates molecular environment information. EnviroDetaNet employs an E(3)-equivariant message-passing neural network combining intrinsic atomic properties, spatial features, and environmental information, allowing it to comprehensively capture both local and global molecular information. Compared to state-of-the-art machine learning models, EnviroDetaNet excels in various predictive tasks and maintains high accuracy even with a 50% reduction in training data, demonstrating strong generalization capabilities. Ablation studies confirm that molecular environment information is crucial for improving model stability and accuracy. EnviroDetaNet also shows outstanding performance in spectral predictions for complex molecular systems, making it a powerful tool for accelerating molecular discovery. |
|---|---|
| ISSN: | 2057-3960 |