A light-weight Graph Neural Network for the prediction of 31P Nuclear Magnetic Resonance signals

Graph Neural Networks (GNNs) are powerful tools for predicting chemical shifts in Nuclear Magnetic Resonance (NMR) spectroscopy. In this paper, we improve the state-of-the-art mean absolute error (MAE) on the Ilm-NMR-P31 dataset for the prediction of ³¹P shifts from 11.4 ppm to 8.88 ppm by proposing a lightweight GNN which is based on the Metalayer-Framework. Furthermore, we analyze the performance of our model depending on the size of the training dataset and compare our model with different state-of-the-art models. Finally, we demonstrate how the GNN's predictions can be interpreted and visualized with respect to the underlying molecular structures. Using GNNExplainer, we analyze the best- and worst-predicted molecules and perform feature ablations to assess the model's reliance on specific input features. Furthermore, we validate the model's physical plausibility by extracting learned substituent effects: the GNN autonomously rediscovers established empirical rules, accurately reproducing the <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mi>β</mml:mi></mml:math>-deshielding and <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mi>γ</mml:mi></mml:math>-shielding increments of neutral phosphines. Finally, we show that the SDGNN outperforms standard vacuum DFT calculations and HOSE code baselines.Scientific Contribution: The proposed lightweight GNN based on the Metalayer framework improves the state-of-the-art ³¹P NMR shift prediction. By systematically varying the training-set size and benchmarking against multiple state-of-the-art models, we provide a standardized performance comparison that was previously lacking for ³¹P NMR shift prediction. Using GNNExplainer and targeted feature ablations, we relate the model's predictions to specific molecular substructures and input features. By quantitatively verifying that the model learns fundamental physical trends like substituent increments and identifying specific error sources, we provide a level of chemical interpretability and validation that goes beyond prior black-box approaches.