Evaluation of Generalized Born Models for Large Scale Affinity Prediction of Cyclodextrin Host-Guest Complexes

Binding affinity prediction with implicit solvent models remains a challenge in virtual screening for drug discovery. In order to assess the predictive power of implicit solvent models in docking techniques with Amber scoring, three generalized Born models (GB(HCT), GB(OBC)I, and GB(OBC)II) available in Dock 6.7 were utilized, for determining the binding affinity of a large set of beta-cydodextrin complexes with 75 neutral guest molecules. The results were compared to potential of mean force (PMF) free energy calculations with four GB models (GB(Still), GB(HCT), GB(OBC)I, and GB(OBC)II and to experimental data. Docking results yield similar accuracy to the computationally demanding PMF method with umbrella sampling. Neither docking nor PMF calculations reproduce the experimental binding affinities, however, as indicated by a small Spearman rank order coefficient (similar to 0.5). The binding energies obtained from GB models were decomposed further into individual contributions of the binding partners and solvent environments and compared to explicit solvent simulations for five complexes allowing for rationalizing the difference between explicit and implicit solvent models. An important observation is that the explicit solvent screens the interaction between host and guest much stronger than GB models. In contrast, the screening in GB models is too strong in solutes, leading to overestimation of short-range interactions and too strong binding. It is difficult to envision a way of overcoming these two opposite effects.