Abstract
Understanding the interactions that make up protein-protein or protein-peptide interfaces is a crucial step towards applications in biotechnology. The mutation of a single residue can have a strong impact on binding affinity and specificity, which is difficult to capture in sampling and scoring. Many established computational methods provide an estimate of binding or non-binding; however, comparing highly similar ligands is an important and significantly more challenging problem. Here we evaluated the capability of predicting ligand binding specificity using three established but conceptually different physics-based methods for protein design. As a model system, we analyzed the binding of peptides to designed armadillo repeat proteins, where a single residue of the peptide was changed systematically, leading to affinity changes in the range of 1-1000 nM. We critically assessed the prediction accuracy of the computational methods. While a good correlation with experimentally determined data was observed in several cases, specific biases in the prediction performance of each method were identified.