Abstract
Photochemically induced dynamic nuclear polarization (photo-CIDNP) is a hyperpolarization method used to boost signal sensitivity in NMR spectroscopy. So far, there is no theory to predict the steady-state photo-CIDNP enhancement reliably, and hence, suitable target molecules need to be identified through tedious experimental screenings. Here, we explore the use of machine learning to predict steady-state photo-CIDNP enhancement. For a series of 27 indole-, five amino-acid-, and eight phenol-derivatives, the signal-to-noise enhancement (SNE) of steady-state photo-CIDNP experiments was measured and then connected to a combination of eight molecular features. The nucleophilic Fukui index was identified as a strong qualitative indicator of the site with the highest SNE in each molecule. Furthermore, a semiquantitative machine learning model based on Logistic Regression identified the sites with high enhancements (SNE > 90) in 100% of cases. Among several quantitative machine learning models for enhancement prediction, CatBoost Regressor and K-Nearest Neighbors showed the best performance. The results demonstrate the high potential of machine learning approaches for predictions of photo-CIDNP SNE, which will enable virtual prescreening of compound libraries.