Predicting molecular structure from sparse NMR data, proof-of-concept using a graph neural network.

We demonstrate that graph-based neural networks are capable of molecular structure elucidation of organic molecules directly from experimentally measurable NMR spectroscopic properties, simulated by DFT. Such a system is a holy grail for researchers in chemical science for molecules in the solution state. Our graph-based transformer neural network architecture is based on 'inverting' IMPRESSION-G2 (which outputs predicted NMR properties based on input molecular structures) to instead output a predicted molecular structure in one shot from input computed NMR properties. The inclusion of scalar (J) coupling input NMR features is shown to be critical to this approach, correctly assigning ∼99.9% of bonds and correctly predicting >95% of molecules in one-shot when all theoretical J values are included (the 'Everything' model). When simulated NMR data are limited to only the experimentally realistic 1H and 13C NMR parameters ('Experimentally measurable simulated parameters' model), the success rate drops substantially to ∼10%. This can be improved by using multi-shot predictions combined with confidence metrics and systematic searching of heteroatom bonding structures to deliver 78.3% success rate on molecules with <2 nitrogens and <2 oxygens and 32.4% overall for the testing set used. This proof-of-concept demonstrates the potential to map spectroscopic properties directly onto molecular structure properties using graph neural networks and highlights key sensitivities to spectroscopic data sparsity and heteroatom degeneracy when taking this approach.