MIM-ML: Combining Molecular Fragmentation and Machine Learning for Accurate Prediction of NMR Chemical Shifts for Large Peptides.

Accurate prediction of NMR chemical shifts for large biomolecules remains a major challenge due to the steep scaling of quantum-mechanical (QM) models with system size. Our group has developed a systematic fragmentation-based approach known as molecules-in-molecules (MIM) that is applicable for large biomolecular systems and can calculate their NMR spectra accurately. Here, we developed a hybrid fragmentation-based machine learning framework (MIM-ML) for predicting 1H and 13C NMR chemical shifts of large peptides. Low-cost single-layer MIM1-based DFT calculations with a microsolvation model were employed to generate the training data. For ML training, physics-based structural and electronic features were incorporated and derived from cost-efficient extended tight-binding (xTB) calculations. Our trained XGBR model using a relatively small amount of data (1406 for 1H and 837 for 13C) achieved mean absolute deviations (MADs) of 0.33 ppm for 1H and 2.91 ppm for 13C against experimental data for a test data set of large peptides. Furthermore, training the ML model directly on experimental chemical shifts produced results consistent with those of the DFT-trained model, indicating independence from the source of training data and a strong correlation between DFT-computed and experimental values. Overall, our developed ML model can be applied for the accurate prediction of the NMR chemical shifts for unknown large peptides.