Acidity Prediction in Arbitrary Solvents: Machine Learning Based on Semiempirical Molecular Orbital Calculation

Due to the nonlinearity of solvent effects, careful solvent selection is essential when using acids in different applications. However, there is a lack of measurements of p<i>K</i>_a while systematically changing molecular structures and solvents. Consequently, there was no protocol to predict the acidity in arbitrary environments. This study developed an arbitrary environment p<i>K</i>_a prediction protocol by integrating quantum chemical calculations using a polarizable continuum model and machine learning. This protocol constructed models to predict the acidity of biologically relevant molecules in water and candidate superstrong acids in organic solvents. For both systems, the p<i>K</i>_a can be predicted with an average absolute error of 1.1 by learning relatively small number of data. This approach can also account for the nonlinear decay of acidity with solvation in different environments (compression effect). The versatility of the protocol extends to its applicability to a wide range of compounds, including those with complex electronic state changes upon proton dissociation, supporting research in diverse fields including, but not limited to, drug discovery and engineering.