Computational modeling of peptide-protein binding

Peptides offer great promise as targeted affinity ligands, but the space of possible peptide sequences is vast, making experimental identification of lead candidates expensive, difficult, and uncertain. Computational modeling can narrow the search by estimating the affinity and specificity of a given peptide in relation to a predetermined protein target. The predictive performance of computational models of interactions of intermediate-length peptides with proteins can be improved by taking into account the stochastic nature of the encounter and binding dynamics. A theoretical case is made for the hypothesis that, because of the flexibility of the peptide and the structural complexity of the target protein, interactions are best characterized by an ensemble of possible bound configurations rather than a single “lock and key” fit. A model incorporating these factors is proposed and evaluated. A comprehensive dataset of 3,924 peptide-protein interface structures was extracted from the Protein Data Bank (PDB) and descriptors were computed characterizing the geometry and energetics of each interface. The characteristics of these interfaces are shown to be generally consistent with the proposed model, and heuristics for design and selection of peptide ligands are derived. The curated and energy-minimized interface structure dataset and a relational database containing the detailed results of analysis and energy modeling are made publicly available via a web repository. A novel analytical technique based on the proposed theoretical model, Virtual Scanning Probe Mapping (VSPM), is implemented in software to analyze the interaction between a target protein of known structure and a peptide of specified sequence, producing a spatial map indicating the most likely peptide binding regions on the protein target. The resulting predictions are shown to be superior to those of two other published methods, and support the validity of the stochastic binding model.