Generalizations in Docking Calculations
Docking a ligand into a protein binding site is a challenging task, fraught with many sources of error. While computing accurate binding affinities for protein-ligand pairs would be exceedingly useful, obstacles persist. Calculating the interaction energy between a flexible protein and a ligand is particularly difficult. Given the many practical obstacles, well-placed approximations and generalizations might, and often do, provide answers to the simpler question of which ligands bind better than others.
A recent study supports the thoughtful use of some surprisingly broad generalizations. Rocky and Elcock (J. Med. Chem. 2005, 48, 4138-4152) predicted relative affinities of an inhibitor for a family of homologous receptors. They examined the binding of seven protein kinase inhibitors to panels of about 20 kinases using two major assumptions. First, given one receptor-ligand complex, they modeled the other receptors with the same backbone conformation, optimizing side chains only. Second, the inhibitor remained fixed; attempts to optimize the inhibitor in the binding sites of the modeled kinases gave inferior results. They used a simple energy function to predict relative binding affinities and were able to reproduce experimental trends fairly well.
The fixed-inhibitor and fixed-backbone assumptions may seem like a stretch at first glance, though it may work here because of some peculiarities of kinase-inhibitor binding. Certain interactions are conserved nearly across the board, and the assumptions ensure that these contributions are equivalent in all cases. For example, hydrogen bonds between the protein backbone and inhibitor are scored in exactly the same way for every pair. Given the many sources of error in any docking calculation, which are even more amplified in cases involving homology models, allowing the inhibitor-backbone interactions to change freely on a case by case basis probably adds more noise than information. The goal in this case was not to find the most accurate binding pose, but rather to determine which kinases bind better than others, and the assumptions work surprisingly well.
An analogy can be made to 3D-QSAR using CoMFA, which requires molecules in the training set to be aligned. Alignments based on conformations of the ligands in the protein binding site typically give inferior results to alignments based on the ligand structures alone. While it’s more realistic to assume similar molecules shift slightly in the binding pocket, alignments that reflect small shifts add noise to the calculation. The work of Rockey and Elcock emphasizes the benefit of averaging out noise in complex calculations.
Jennifer Shepphird, IRC
1. W.M. Rockey & A.H. Elcock. Rapid Computational Identification of the Targets of Protein Kinase Inhibitors.
Leave Your Comments ...
You must be logged in to post a comment.