Protein–Ligand Complex Solvation Thermodynamics: Development, Parameterization, and Testing of GIST-Based Solvent Functionals
datasetposted on 22.01.2020, 20:21 by Tobias Hüfner-Wulsdorf, Gerhard Klebe
In drug design, the importance of molecular solvation and desolvation is increasingly appreciated and water molecules are recognized as active contributors to protein–ligand binding. However, despite a number of theoretical approaches, computational tools are still far from routinely integrating solvation features into rational structure–affinity relationships (SARs). In this contribution, we present a set of solvent functional-based models, which calculate the relative binding free energy contributions resulting from solvation for a diverse set of 53 thrombin protein–ligand complexes. These protein–ligand complexes were further matched into chemically similar pairs of ligand molecules. Our solvent functionals are based on molecular dynamics simulations in conjunction with grid inhomogeneous solvation theory (GIST) processing, and they are calibrated using accurate experimental data from isothermal titration calorimetry (ITC) measurements. We found that excellent agreement with experimental measurements can be achieved by considering either the desolvation of the protein-binding pocket or the ligand in solution prior to binding. The incorporation of contributions from the protein–ligand complexes generally results in good agreement with experimental measurements but require additional adjustment of spatial cutoff parameters. In addition, we investigated the transfer of the trained solvent functionals to another protein target, which revealed deviating performance results, indicating a target-specific treatment of solvation features within the model. Together with our tool GIST-based processing of solvent functionals (Gips), we provide a way to automatically generate solvent functional parameters from GIST data and allow for the design of compounds with favorable solvation properties given the chemical similarity and affinity range of the matching pairs in the training set. Finally, we reflect on the resemblance with the popular three-dimensional quantitative SAR (3D-QSAR) method, as our study allows for (retrospective) insights on the high predictive power of this well-established method.