pKa prediction for small molecules: an overview of experimental, quantum, and machine learning-based approaches

The pKa, also known as the logarithmic dissociation constant, is a crucial parameter that defines the ionization level of a molecule when it is in solution. It is essential for several physicochemical properties, including lipophilicity, solubility, protein binding affinity, and the ability to cross biological membranes. Therefore, obtaining accurate pKa assessments is vital for modifying and refining the acidity and basicity of organic compounds. Accurate prediction can help improve drug design, optimize pharmaceutical formulations, analyze the behavior of pollutants in the environment, and guide the development of new materials. Traditionally, pKa determination has relied on experimental techniques. However, the recent emergence of machine learning (ML) has led to significant advances in pKa prediction. In this review, we examine various approaches for pKa prediction, with a focus on recent advances in machine learning. We discuss the performance of these models, drawing on results reported in publications related to the SAMPL Challenges and Novartis prediction challenges. Because of their different theoretical and computational frameworks, protein pKa prediction methods are not included in this review, which focuses exclusively on small organic molecules. Finally, we highlight current challenges and future directions, including the integration of hybrid models combining quantum mechanics and machine learning, the improvement of benchmark databases, and the development of more universal and interpretable predictive models. We hope that this paper can provide useful guidelines for future research.