Graph Convolutional Neural Networks (GCNNs) for accurate NMR ¹⁹F-NMR chemical shift prediction

Accurate prediction of NMR chemical shifts is the foundation for spectral assignment, structural elucidation and automated structural verification. In the work reported here, we are developing Graph Convolutional Neural Network (GCNNs) to represent molecular structures and learn their relationships with ¹⁹F-NMR chemical shifts. In this framework, molecules are treated as graphs in which atoms are nodes and bonds are edges. Node features such as valency, electronegativity, and hybridisation capture each atom’s chemical environment, allowing the network to infer relationships between structure and NMR shift directly from data. 

The GCNN exploits ‘message passing’, where each node exchanges information with neighbouring atoms to learn its extended chemical context, followed by prediction, where aggregated node features are processed through a neural network trained on experimental chemical shifts. Such networks are referred to as Message Passing neural Networks (MPNNs). The approach captures subtle dependencies between molecular structure and NMR properties that conventional empirical methods can miss.