Computational Thrombin Inhibitor Optimization

Abstract

Of the plethora of computational methods for the estimation of binding free energies, molecular dynamics based alchemical approaches provide remarkable accuracy at an affordable computational cost. The computationally less expensive empirical ligand and structure based methods require training a predictive model to yield accurate free energy estimates. In the current study we utilized a combinatorial chemistry library and combined the chemoinformatic and alchemical approaches to optimize a set of known thrombin inhibitors. A confirmed ligand scaffold (D-Phe-Pro) was selected as a basis for modifications. Prior to designing new compounds, we verified the non-equilibrium alchemical free energy calculation setup on a set of inhibitors with experimentally measured binding affinities. Afterwards, a combinatorial chemistry library was constructed by modifying substituents at the ortho, meta and para positions on a benzene ring facing the S1 thrombin pocket. Navigation in the library was guided by training the Partial Least Squares based regression models using topological, structural, physicochemical descriptors and COMFA-like coordinates of the molecular electrostatic surfaces. The predicted potential binders were subjected to the alchemical free energy calculations, and, subsequently, the obtained estimates were used to update the regression models and improve prediction accuracy. Few iterations sufficed to reach convergence towards stronger binders, several of which were predicted to have a higher binding affinity than the strongest inhibitors in the initial compound set. While the newly identified ligands are to be validated in an ITC experiment, the computational workflow utilizing chemoinformatic approaches for navigation in a combinatorial chemistry library and first principles based alchemical calculations for accurate free energy estimates appears to be a powerful approach for ligand optimization.