Solvation Free Energies of Anions: From Curated Reference Data to Predictive Models

Area of Science:

• Computational Chemistry
• Physical Chemistry
• Machine Learning in Chemistry

Background:

• Accurate prediction of physicochemical properties (solubility, lipophilicity) is crucial for ionizable solutes.
• Reliable prediction methods require high-quality reference data for solvation free energies of ions, which is currently limited.
• Existing computational methods face challenges due to data scarcity and accuracy.

Purpose of the Study:

• To address the limitations in data quality and availability for ionic property prediction.
• To develop accurate and inexpensive computational models for predicting pKa, gas-phase acidities, and solvation free energies of anions.
• To provide publicly accessible databases and machine learning models for ionic phenomena.

Main Methods:

• Compiled and curated databases for pKa (8,241 points, 8 solvents) and gas-phase acidities (5,536 points) using QM calculations.
• Calculated solvation free energies for anions (6,090) and neutral conjugate solutes (6,088) using thermodynamic cycles and COSMO-RS.
• Trained graph neural network (GNN) models using reaction SMILES for pKa/gas-phase acidity and SMILES for anions to predict solvation energies.

Main Results:

• The microscopic pKa model achieved high accuracy (0.58-0.59 MAE) on unseen data and the SAMPL7 challenge.
• Gas-phase acidity models showed mean absolute errors slightly above 2 kcal/mol against experimental data.
• Anionic solvation free energy models demonstrated mean absolute errors below 3 kcal/mol, comparable to QM-based methods.

Conclusions:

• The developed GNN models offer a fast and accurate alternative to traditional QM approaches for predicting ionic properties.
• The publicly available databases and models significantly enhance data availability and quality for ionic phenomena research.
• These advancements facilitate more reliable predictions of essential physicochemical properties for ionizable solutes.