Physics-Informed Machine Learning Enables Rapid Macroscopic pKa Prediction

Accurate prediction of macroscopic pKa values remains a central challenge in computational chemistry, critical for modeling pH-dependent properties like solubility, membrane permeability, and charge state. Here we introduce Starling, a physics-informed neural network based on the Uni-pKa architecture trained to predict per-microstate free energies and compute macroscopic pKavalues via thermodynamic ensemble modeling. Unlike approaches that treat protonation events in isolation, Starling explicitly resolves protonation and tautomeric microstates, enabling robust handling of complex molecules with multiple ionizable sites. We show that Starling achieves comparable or superior accuracy to leading commercial tools on multiple benchmark datasets, and demonstrate its utility in predicting isoelectric points, logD profiles, and blood–brain-barrier permeability. By maintaining thermodynamic consistency and enabling rapid microstate-ensemble generation, Starling enables accurate physicochemical property prediction with broad relevance to drug discovery and molecular design.