Site-Specific pK_a Prediction: Metrologically Validated Hydrogen-Centered Graph Machines

The acid dissociation constant (pK_a) is a fundamental property reflecting the local propensity of a chemical species to release a proton. While graph neural networks excel at predicting global properties, accurate pK_a prediction requires a shift toward atomic mapping. We introduce hydrogen-centered graph machines (HGM), specifically designed for pK_a prediction, by incorporating within the graph the mobile proton involved in the acid-base equilibrium. To support this approach, we refined over 2,300 metrologically validated data points to guarantee fidelity to aqueous species. By calibrating the parametric complexity of HGM across an ensemble of 50 models, we ensured generalization capabilities for robust predictions. As a result, HGM achieved root mean square errors (RMSEs) between 0.6 and 0.7 pK_a units on the SAMPL6, Jensen, and EuroSAMPL1 test sets. However, 2D topological representations remain localized and thus fail to capture charge delocalization—where present—across an entire chemical species, particularly within complex poly-nitrogen systems. To bridge this gap, we introduce a mesomeric weighting strategy that complements the initial structure with relevant resonance forms. This approach significantly enhances predictive accuracy, driving the RMSE down to 0.5 pK_a units for the more complex SAMPL6 and EuroSAMPL1 challenges. Beyond proven accuracy, HGM acts as an automated metrological filter to establish an applicability domain. Thus, HGM identifies structural anomalies—such as complex tautomers and 3D interactions like charge-assisted hydrogen bonds—via statistical warnings that confirm their pK_a prediction falls beyond the actual scope of the HGM framework. As a definitive proof of concept, we deployed HGM to investigate four complex therapeutics (prednisolone, amoxicillin, azithromycin, and pristinamycin I_A). We successfully pinpointed the exact ionization sites and attributed the measured pK_a values to their corresponding equilibria. Benchmarking against the eight available experimental pK_a values yields a mean absolute error of 0.4 pK_a units. Fundamentally, for such complex systems, it is the synergy between the mapping of multiple speciation pathways and the chemist's critical evaluation that resolves the attribution ambiguities inherent to standard commercial databases. We release a standalone Docker container for result verification and out-of-the-box predictions of new compounds.