Uni-pKa: An Accurate and Physically Consistent pKa Prediction through Protonation Ensemble Modeling

Predicting p K a values of small molecules has key applications in drug discovery and molecular simulation. However, current methods face challenges in rigorously interpreting experimental data and ensuring thermodynamic consistency between successive p K a values. To address these limitations, we present Uni-p K a , an accurate and reliable p K a prediction framework. Uni-p K a is based on comprehensive free energy modeling of possible molecules in protonation equilibrium. Within this framework, a structural enumerator recovers underlying structures in p K a datapoints, and a neural network serves as a free energy predictor, learning from data rigorously while inherently preserving thermodynamic consistency. Through a pretraining-finetuning strategy utilizing predicted and experimental p K a data, Uni-p K a achieves state-of-the-art accuracy among chemoinformatic methods. Uni-p K a provides a good example of combining chemical principles and machine learning to solve scientific problems.