Steering protein-ligand docking with quantitative NMR chemical shift perturbations

Lead optimization benefits from including structural knowledge of the target. We present a new method that exploits quantitatively NMR amide proton chemical shift perturbations (CSP) on the protein side for protein-ligand docking. The approach is based on a hybrid scoring scheme consisting of a weighted sum of DrugScore, describing protein-ligand interactions, and Kendall's rank correlation coefficient, which scores ligand poses with respect to their agreement with experimental CSP data. For back-calculating CSP for a ligand pose, an efficient empirical model considering only ring-current effects is applied. The hybrid scoring scheme has been implemented in AutoDock. Compared to previous approaches, the rank correlation provides a measure that is more robust against the presence of outliers in back-calculated CSP data. Furthermore, our methods exploit CSP information at docking time and not for postfiltering, resulting in an enhanced generation of native-like solutions. As we exploit CSP information quantitatively, the experimental information effectively contributes to orient the ligand in the binding site. When applied to 70 protein-ligand complexes with computed CSP reference data, the docking success rate increases from 71%, if no CSP information is used, to 99% at the highest CSP weighting factor tested. Global optimization, thus, performs satisfactorily on the hybrid docking energy landscape. We next applied the approach to three test cases with experimental CSP reference data. Without CSP information, neither of the complexes is successfully docked. Including CSP information with the same CSP weighting factor, as determined above, leads to successful docking in all three cases. Only then native-like ligand configurations are generated at two of the three complexes. Binding site movements of up to 2 A are found to not deteriorate the docking success. The approach will be particularly important for protein-ligand complexes that are difficult to predict computationally, such as ligands binding to flat interface regions.