Prediction of 13C-NMR Chemical Shifts of Small Organic and Drug Molecules Using Machine Learning

Surajit Das (Tata Institute of Fundamental Research Hyderabad, India)

LinkedIn: Surajit Das, X: surajit_das_x, BlueSky: surajitdas.bsky.social

Abstract: Accurate prediction of NMR chemical shifts is essential for mapping the structure-property relationships of complex molecules. In this work, we use the kernel ridge regression (KRR) machine learning architecture with an in-house-built aBoB-RBF(n) atomic representation to predict 13C-NMR chemical shifts. To train the model, we use the QM9NMR dataset, which contains 130,831 molecules with up to 9 non-hydrogen (CONF) atoms, encompassing nearly 832k carbon atomic environments. This vast pool of chemical diversity allows us to train on 100k data points with a 50k test set, yielding a mean absolute error of 1.69 ppm, an improvement over previously reported models. We also demonstrate model performance on diverse validation sets, including the drug-like Drug12 and Drug40 datasets; the GDBm dataset, which contains an increasing number of CONF atoms; and the Pyrimidinone dataset, which features biologically relevant substituted uracil and pyrimidinone molecules. Additionally, we developed a Python-based mlqm9nmr module, available on GitHub, that allows users to quickly predict chemical shifts using our KRR model. Finally, we developed a web-based interface that uses a given molecule’s SMILES string to predict 13C NMR chemical shifts with the trained KRR model.