Efficient Prediction of Transition-Metal NMR Chemical Shifts Using Machine Learning: Do Two-Dimensional Descriptors Suffice?

Accurate prediction of nuclear magnetic resonance chemical shifts for transition-metal nuclei remains a challenging problem due to the high computational cost of quantum-chemical methods and the limited availability of experimental data. In this study, machine learning models were developed to predict chemical shifts of coordination compounds containing Mn, Fe, Nb, and Mo, using a curated data set of 1956 experimental measurements. Several approaches were evaluated, including descriptor-based models, graph neural networks, and transformer-based architectures. The Tabular Prior-Data Fitted Network model demonstrated the best performance across all nuclei, with prediction errors corresponding to 4.5–8% of the total chemical shift range. A comparison of molecular representations showed that two-dimensional descriptors provide accuracy comparable to three-dimensional approaches while requiring significantly lower computational cost. Model interpretation based on Shapley additive explanations revealed metal-specific structure–property relationships and enabled identification of the applicability domain. Attempts to construct unified models across different metals did not improve predictive performance, highlighting the element-specific nature of chemical shifts. External validation on an independent data set of 195 Pt complexes confirmed the generalizability of our two-dimensional descriptor-based approach, achieving a holdout MAE of 159 ppm without any 3D structural information. These results demonstrate that machine learning models based on molecular descriptors provide an efficient and reliable alternative to quantum-chemical methods for predicting nuclear magnetic resonance chemical shifts of transition-metal compounds.