Structure-based prediction of protein–protein binding affinity with consideration of allosteric effect

The conformational change upon protein–protein binding is largely ignored for a long time in the affinity prediction community. However, it is widely recognized that allosteric effect does play an important role in biomolecular recognition and association. In this article, we describe a new quantitative structure–activity relationship (QSAR)-based strategy to capture the structural and nonbonding information relating to not only the direct noncovalent interactions between protein binding partners, but also the indirect allosteric effect associated with binding. This method is then employed to quantitatively model and predict the protein–protein binding affinities compiled in a recently published benchmark consisting of 144 functionally diverse protein complexes with their structures available in both bound and unbound states (Kastritis et al. Protein Sci 20:482–491, 2011). With incorporating genetic algorithm and partial least squares regression (GA-PLS) into this method, a significant linear relationship between structural information descriptors and experimentally measured affinities is readily emerged and, on this basis, detailed discussions of physicochemical properties and structural implications underlying protein binding process, as well as the contribution of allosteric effect to the binding are addressed. We also give an empirical estimation of the prediction limit rpred2 = 0.80 for structure-based method used to determine protein–protein binding affinity.