Message Passing Neural Networks for Predicting 1H and 19F Chemical Shifts

We are using advanced artificial intelligence techniques to improve NMR chemical shift prediction, spectral assignment, and automated structural validation in NMR spectroscopy. Graph neural networks (GNNs) are well-suited for molecular representation, treating atoms as nodes and bonds as edges, making them the leading framework for predictive modelling. Their primary strength lies in their simultaneous use of both node feature information and the atom connectivities encoded in the graph adjacency matrix. A specialised class of GNN, the Message Passing Neural Network (MPNN), defines a receptive field around each atom, enabling learning at the molecular substructure level. This is particularly valuable for modelling chemical shifts, which are highly dependent on the local electronic environment of the active nuclei. We are training ensembles of MPNNs from large collections of molecular structures (>10,000s) annotated with experimentally observed chemical shifts. Each structure in the training set is represented by the adjacency matrix of its underlying graph and a set of atomic features, which include stereo bond configurations and chiral centre information derived from three-dimensional graph embeddings. Outcomes from models trained on proton (1H) and fluorine (19F) chemical shifts are reported here. The median absolute prediction errors from unseen test data are approximately 0.08 ppm for ¹H and 2.2 ppm for ¹⁹F. In both cases, the error distributions are smooth, symmetric about zero, and can be accurately modelled using a Gaussian kernel density estimation. This approach, in turn, enables a data-driven, probabilistic method for spectral assignment and structural verification.