Validation of an automated procedure for the prediction of relative free energies of binding on a set of aldose reductase inhibitors

Among the available methods for predicting free energies of binding of ligands to a protein, the molecular mechanicsPoisson–Boltzmann surface area (MM-PBSA) and molecular mechanics generalized Born surface area (MM-GBSA) approacheshave been validated for a relatively limited number of targets and compounds in the training set. Here, we report the results ofan extensive study on a series of 28 inhibitors of aldose reductase with experimentally determined crystal structures and inhibitoryactivities, in which we evaluate the ability of MM-PBSA and MM-GBSA methods in predicting binding free energies using a numberof different simulation conditions. While none of the methods proved able to predict absolute free energies of binding in quantitativeagreement with the experimental values, calculated and experimental free energies of binding were significantly correlated.Comparing the predicted and experimental DG of binding, MM-PBSA proved to perform better than MM-GBSA, and within theMM-PBSA methods, the PBSA of Amber performed similarly to Delphi. In particular, significant relationships between experimentaland computed free energies of binding were obtained using Amber PBSA and structures minimized with a distance-dependentdielectric function. Importantly, while free energy predictions are usually made on large collections of equilibrated structures sampledduring molecular dynamics in water, we have found that a single minimized structure is a reasonable approximation if relativefree energies of binding are to be calculated. This finding is particularly relevant, considering that the generation of equilibrated MDensembles and the subsequent free energy analysis on multiple snapshots is computationally intensive, while the generation andanalysis of a single minimized structure of a protein–ligand complex is relatively fast, and therefore suited for high-throughput virtualscreening studies. At this aim, we have developed an automated workflow that integrates all the necessary steps required togenerate structures and calculate free energies of binding. The procedure is relatively fast and able to screen automatically and iterativelymolecules contained in databases and libraries of compounds. Taken altogether, our results suggest that the workflow can bea valuable tool for ligand identification and optimization, being able to automatically and efficiently refine docking poses, whichsometimes may not be accurate, and rank the compounds based on more accurate scoring functions.