Small organic compounds’ reduction/oxidation (redox) potential is a key property that drives innumerable chemical and biological electron transfer reactions. However, experimental measurement of redox potential is time-consuming and expensive, yielding few and small experimental measured data sets. Computational methods have previously been applied to create redox predictors applicable only to a specific data set. In this work, we investigate the effectiveness of various descriptors, including structural and functional properties, molecular energies, and drug-like properties, to predict redox potential. We use Gaussian Process Regression (GPR) as a model, as it is suitable for fitting small data sets and has shown promise in predicting redox potential. We train and test our redox predictor on three organic molecule data sets with increasing structural diversity. We demonstrate that a GPR-based redox predictor using individual features or a combination of molecular descriptors (DFT energies, molecular descriptors, and ADME properties works well across quinone-based and a general organic data set. Finally, we test the trained model against an experimental data set of quinones to assess if the model makes predictions well correlated with experimental data.
Kalia et al. (Mon,) studied this question.