Improving inverse docking target identification with Z-score selection.

The utilization of inverse docking methods for target identification has been driven by an increasing demand for efficient tools for detecting potential drug side-effects. Despite impressive achievements in the field of inverse docking, identifying true positives from a pool of potential targets still remains challenging. Notably, most of the developed techniques have low accuracies, limit the pool of possible targets that can be investigated or are not easy to use for non-experts due to a lack of available scripts or webserver. Guided by our finding that the absolute docking score was a poor indication of a ligand's protein target, we developed a novel "combined Z-score" method that used a weighted fraction of ligand and receptor-based Z-scores to identify the most likely binding target of a ligand. With our combined Z-score method, an additional 14%, 3.6%, and 6.3% of all ligand-protein pairs of the Astex, DUD, and DUD-E databases, respectively, were correctly predicted compared to a docking score-based selection. The combined Z-score had the highest area under the curve in a ROC curve analysis of all three datasets and the enrichment factor for the top 1% predictions using the combined Z-score analysis was the highest for the Astex and DUD-E datasets. Additionally, we developed a user-friendly python script (compatible with both Python2 and Python3) that enables users to employ the combined Z-score analysis for target identification using a user-defined list of ligands and targets. We are providing this python script and a user tutorial as part of the supplemental information.