Solvation free energies of anions: from curated reference data to predictive models

Predicting the physicochemical properties of ionizable solutes, including solubility and lipophilicity, is of broad significance. Such predictions rely on the accurate determination of solvation free energies for ions. However, the limited availability of high-quality reference data poses a challenge in developing accurate, inexpensive computational prediction methods. In this study, we address both issues of data quality and availability. We present three databases and models related to ionic phenomena: (1) 8,241 pKa data points across 8 solvents, (2) 5,536 gas-phase acidities from DLPNO-CCSD(T) QM calculations, and (3) 6,090 solvation free energies of anions across 8 solvents obtained from a thermodynamic cycle. We also report 6,088 solvation free energies of neutral conjugate solutes computed using the COSMO-RS method. The pKa data were obtained from the iBonD database, cleaned, and combined with a separate compilation of trustworthy reference pKa data. Gas-phase acidities were computed for most of the acids present in the pKa corpus. Leveraging these data, we compiled values for solvation free energies of anions. We then trained several graph neural network models, which can be used as an alternative to QM approaches to quickly estimate these properties. The pKa and gas-phase acidity models accept reaction SMILES strings of the acid dissociation as inputs, whereas the solvation energy model accepts the SMILES string of the anion. Our microscopic pKa model achieves good accuracy, with an overall test mean average error of 0.58 units on unseen solutes and 0.59 on the SAMPL7 challenge (the lowest error so far among multisolvent models). Our gas-phase acidity model had mean absolute errors slightly above 2 kcal mol-1 when evaluated against experimental data. The anionic solvation free energy model had mean absolute errors of less than 3 kcal mol-1 in several test evaluations, comparable to (though less reliable than) several widely used QM-based solvation models.