CSP5: Large-scale Neural Chemical Shift Prediction from 2.5 Million Experimental NMR Spectra

Accurate nuclear magnetic resonance (NMR) chemical shift prediction remains limited by the small size and narrow chemical diversity of atom-assigned training datasets. Here, we address that bottleneck by leveraging the new NMRexp resource with weak supervision, training CASCADE-style graph neural networks with an optimal-assignment loss with datasets of approximately 1.2 million unassigned 13C spectra and 1.4 million unassigned 1H spectra alongside conventional assigned data. This joint training strategy substantially improves performance over assigned-only baselines: on held-out NMRexp data, 13C mean absolute error (MAE) falls from 1.46 to 1.05 ppm and 1H MAE from 0.316 to 0.134 ppm, while 13C performance on a test set from the assigned Exp22K dataset improves from 0.812 to 0.610 ppm. These gains also translate into better structure verification performance, with hard-decoy top-1 retrieval improving from 0.819 to 0.876 for 13C and from 0.589 to 0.801 for 1H. Taken together, these results show that large-scale unassigned literature spectra provide an effective training signal for more accurate and practically useful NMR prediction models.