• Protein[BOND]peptide interaction;
  • Knowledge-based statistical potential;
  • Quantitative structure-activity relationship;
  • Binding affinity


Protein[BOND]peptide interactions have recently been found to play an essential role in constructing intracellular signaling networks. Understanding the molecular mechanism of such interactions and identification of the interacting partners would be of great value for developing peptide therapeutics against many severe diseases such as cancer. In this study, we describe a structure-based, general-purpose strategy for fast and reliably predicting protein[BOND]peptide binding affinities. This strategy combines unsupervised knowledge-based statistical potential derived from 505 interfacially diverse, non-redundant protein[BOND]peptide complex structures and supervised quantitative structure-activity relationship (QSAR) modeling trained by 250 protein[BOND]peptide interactions with known structure and affinity data. The built partial least squares (PLS) model is confirmed to have high stability and predictive power by using internal 5-fold cross-validation and rigorous Monte Carlo cross-validation (MCCV). The model is further employed to analyze two large groups of HLA- and SH3-binding peptides based upon computationally modeled structures. Satisfactorily, although the PLS model is originally trained with dissociation constants (Kd) of protein[BOND]peptide binding, it shows a good correlation with other two affinity qualities, i.e. SPOT signal intensities (BLU) and half maximal competitive concentrations (IC50). Furthermore, we perform systematic comparisons of our method with several widely used, representative affinity predictors, including molecular mechanics-based MM-PB/SA, knowledge-based DFIRE and docking score HADDOCK, on a small panel of elaborately selected protein[BOND]peptide systems. It is demonstrated that (i) the QSAR-improved statistical potential exhibits a comparable predictive performance with but can work faster than these traditional methods, and (ii) the crystal structure-derived statistical potential also supports the modeled and solution structures of protein[BOND]peptide complexes. We expect that this hybrid method can be exploited as a new scoring tool to facilitate, for example, peptide docking and virtual screening.