| Résumé | The present study provides a comprehensive systematic analysis on the applicability of the linear interaction energy (LIE) approximation to the prediction of gas-to-water transfer (hydration) free energy. The study is based on molecular dynamics simulations in explicit solvent for an extensive and diverse hydration data set comprising 564 neutral compounds with measured hydration free energies, including a “traditional” data set and the more challenging drug-like SAMPL1 data set. A highly correlative LIE model was achieved without empirical scaling of the solute-solvent interaction energy terms along with a cavity term calibrated to the experiment. This model was particularly accurate for the “traditional” data set and of acceptable accuracy for the SAMPL1 data set, with mean-unsigned-errors below 1 kcal/mol and slightly above 2 kcal/mol, respectively. We have analyzed the sensitivity of the LIE model to several parameters such as continuum correction terms applied outside the explicit water shell, the impact of various charging methods, the applicability of single-conformer representation of the solute, and the inclusion of internal energy terms. The parameters with the greatest sensitivity are the charging methods used, with AM1BCC-SP (without AM1 geometry optimization) charges favored over AM1BCC-OPT and RESP charges. The inclusion of the change in intramolecular van der Waals and electrostatic energies between the solution and gas phases can also lead to improved prediction accuracies. Functional group based error analysis identified several chemical classes as minor outliers with systematic errors. A direct comparison of the LIE and free energy perturbation (FEP) approaches using the same force field and charging method shows that the LIE approximation is at least as accurate as the FEP approach with a reduction of computing time by at least 1 order of magnitude. |
|---|
