Predicting scalar coupling constants by graph angle-attention neural network.

Area of Science:

• Computational Chemistry
• Artificial Intelligence
• Spectroscopy

Background:

• Scalar coupling constants (SCCs) from nuclear magnetic resonance (NMR) are vital for molecular structure analysis.
• Experimental NMR and computational density functional theory (DFT) methods for SCC determination are costly and time-consuming.
• Graph neural networks (GNNs) offer potential for rapid, data-driven SCC prediction.

Purpose of the Study:

• To develop an efficient and accurate AI model for predicting SCCs.
• To leverage easily accessible molecular information, including geometric angles, for SCC prediction.
• To enhance the physicochemical interpretability of AI models in molecular property prediction.

Main Methods:

• Proposed a Graph Angle-Attention Neural Network (GAANN) model incorporating a multilayer message-passing network and self-attention.
• Integrated a priori knowledge of molecular angles (bond angles, dihedral angles) into the AI model.
• Combined AI with quantum chemistry theory (Karplus equation) for physicochemical interpretability.

Main Results:

• GAANN accurately simulates molecular topology and predicts SCCs with a log(MAE) of -2.52, comparable to DFT.
• The model demonstrates strong physicochemical interpretability regarding the influence of angles on SCCs.
• Analysis revealed bond angles have the highest correlation with SCCs among various angle features in small molecules.

Conclusions:

• GAANN provides a computationally efficient and accurate alternative for SCC prediction.
• The model's interpretability enhances understanding of the relationship between molecular geometry and SCCs.
• This approach facilitates large-scale SCC analysis for unknown molecules, advancing molecular structure determination.