MolNet_Equi: A Chemically Intuitive, Rotation-Equivariant Graph Neural Network.

Area of Science:

• Computational Chemistry
• Machine Learning
• Deep Learning

Background:

• Deep learning (DL) models show promise for chemical predictions but are limited to scalar properties.
• Predicting vectorial properties like molecular dipole moments requires handling 3D spatial orientation.

Purpose of the Study:

• To develop a rotation-equivariant deep learning model (MolNet_Equi) for accurate prediction of molecular properties.
• To enable DL models to understand and predict directional properties, such as molecular dipole moments, in 3D space.

Main Methods:

• Developed MolNet_Equi, a rotation-equivariant model that recognizes molecular rotation in 3D Euclidean space.
• Ensured rotational equivariance by designing operations where rotation, prediction, and inverse-rotation do not alter the dipole moment prediction.
• Trained the model using dipole-moment magnitudes while enabling prediction of absolute dipole moment directions.

Main Results:

• MolNet_Equi demonstrates superior performance in predicting scalar properties.
• The model accurately predicts directional dipole moments in a rotation-sensitive manner.
• MolNet_Equi successfully predicts absolute dipole moment directions from molecular poses, even when trained only on magnitudes.

Conclusions:

• Incorporating fundamental chemical principles into DL models leads to more chemically intuitive and powerful tools.
• MolNet_Equi represents a significant advancement in applying deep learning to complex chemical property prediction.
• This approach opens new avenues for DL in understanding and predicting vectorial molecular properties.