Multistage Autotuning Pipeline using Simulated Annealing for Parametric Optimization of the Langevin Dipole Solvation Model
Solubility is a critical attribute of pharmacologically relevant compounds.Predictive computational models of solvation free energy can be used as an effective discriminator for screening libraries in the early stages of drug discovery and development. The Langevin Dipole (LD) solvent model is a consistent 3D grid-based approach that takes into account the effective partial atomic charges of molecules. This study focuses on the implementation and optimization of the commercially available OPLS forcefield charges into the LD model for the accurate evaluation of solvation free energy. The Minnesota Solvation Database was used as our benchmark. Key tuning parameters include grid cell spacing and van der Waals radii for various atom types.We optimize the predictive accuracy by systematically tuning these parameters in multiple stages, focusing on one class of compounds per stage. Separating the optimization into stages allows for better traceability and reduces spillover of error between compound classes. In each stage, optimal radii were obtained using simulated annealing. Overall, our results demonstrate a significant improvement in the predictive power of our optimized LD model over commercially available Generalized Born solvation model.