Structure-Based Modelling of Solubility Parameters, Addressing Challenges in Predicting Solvation Energy for Enhanced Model Performance with GBSA

Abstract

Generalised Born models offer a rapid and accurate approach for predicting solvation energies, presenting a viable alternative for molecular solvation predictions. This project aimed to develop a simplified, minimally parameterised Generalised Born model to streamline its implementation and usability. Refinements were made by reparameterising radial shifts and introducing a hydrogen bonding correction within the GBSA framework. This involved optimising solvation energy components using AIMD calculations for Gshift and leveraging solutes from the Minnesota Solvation Database for the electrostatic component.
The parameterisation process focussed on neutral solutes across 11 solvents, comprising nine nonpolar and two polar solvents. Classification of the solutes revealed varying model performance, with aliphatic amines outperforming other nitrogen-containing solutes. A hydrogen bonding correction model, which was particularly effective for water and octanol solvents, was implemented and incorporated atom charge and solvent-accessible surface area contributions. Validation with the Freesolv database demonstrated significant improvements in model performance, yielding a RMSE of 2.28 kcalmol−1 and a MAE of 1.78 kcalmol−1. This performance is competitive with the SM12 model (MAE = 0.58 kcalmol−1 ), as well as with the PC-SAFT model (MAE = 0.58 kcalmol−1 ), an equation of state based on statistical associating fluid theory (SAFT). While these models exhibit lower MAE values, they were trained using much higher levels of theory with larger basis sets. The incorporation of the hydrogen bonding correction model further refined the predictions of GBSA-M1, demonstrating superior performance in predicting pKa values (MAE = 1.11) for small organic monoprotic compounds in water. This represents a significant improvement over established models, which can exhibit prediction errors of up to 7 pKa units, further validating the model’s efficacy and broad applicability.

In a separate study, the effectiveness of structure-based models for predicting Hansen Solubility Parameters (HSPs) of Conjugated Microporous Polyanilines (CMPAs) was assessed using LOOCV with a random forest regressor. However, the models exhibited inadequate alignment with experimental data, highlighting significant flaws in the data quality that compromised the model’s predictive accuracy. Challenges in accurately modelling CMPAs due to potential solubility issues prompted a secondary approach, focusing on higher molecular weight CMPAs to capture complex intermolecular interactions. While this secondary model showed improvement for certain parameters, such as δD, challenges remained, suggesting the need for further refinement and investigation into the underlying data integrity issues.

Date of Award	4 Feb 2025
Original language	English
Awarding Institution	University of Bristol
Supervisor	Natalie Fey (Supervisor) & Frederick R Manby (Supervisor)