Experimental versus predicted affinities for ligand binding to estrogen receptor: iterative selection and rescoring of docked poses systematically improves the correlation.
The computational determination of binding modes for a ligand into a protein receptor is much more successful than the prediction of relative binding affinities (RBAs) for a set of ligands. Here we consider the binding of a set of 26 synthetic A-CD ligands into the estrogen receptor ERα. We show that the MOE default scoring function (London dG) used to rank the docked poses leads to a negligible correlation with experimental RBAs. However, switching to an energy-based scoring function, using a multiple linear regression to fit experimental RBAs, selecting top-ranked poses and then iteratively repeating this process leads to exponential convergence in 4-7 iterations and a very strong correlation. The method is robust, as shown by various validation tests. This approach may be of general use in improving the quality of predicted binding affinities.