Automated Large-Scale File Preparation and Docking: Evaluation of ITScore and STScore Using the 2012 Community Structure−Activity Resource Benchmark

We present the first study utilizing the full set of compounds from the recently released 2012 Community Structure−Activity Resource (CSAR) data set. The CSAR data set is a realistic benchmark for protein-ligand docking scoring functions, containing 57 crystal structures and 757 compounds, most with known affinities from pharmaceutical companies. We used the CSAR data set to evaluate two knowledge-based scoring functions, ITScore and STScore, and a simple force-field-based potential, FFPScore. To conduct this large-scale docking evaluation, we scripted our docking software and associated tools for automated preparation, docking, and evaluation, enabling others to reproduce our results. We also developed a new tool for standardizing the order and name of atoms within MOL2 files. The full set of compounds was used to evaluate binding affinity predictions and active/inactive compounds discrimination. The subset that includes crystal structures was used, in addition, to evaluate the binding mode prediction accuracy of the scoring functions. Within this structure subset, we investigated the importance of accurate ligand and protein conformational sampling and found that the binding affinity predictions are less sensitive to non-native ligand and protein conformations than the binding mode predictions. The full CSAR data set was found to be more challenging in making binding affinity predictions than the subset with structures. Our study suggests the importance of sampling the ligand conformations simultaneously with energy scoring, in contrast to the popular practice of docking pre-generating ligand conformations as rigid bodies. The CSAR docking preparation scripts are offered freely to the academic community.