Assessing the accuracy of binding pose prediction for kinase proteins and 7-azaindole inhibitors: a study with AutoDock4, Vina, DOCK 6, and GNINA 1.0

Abstract

Comparative benchmarking of molecular docking tools is vital for assessing the reliability of virtual screening and binding pose prediction in drug discovery. Our study evaluates the performance of four open-source docking programs: AutoDock4, AutoDock Vina, DOCK 6, and GNINA 1.0 in predicting binding poses of 7-azaindole derivative compounds across 70 kinase–ligand complexes from the RCSB Protein Data Bank. These compounds were selected due to the azaindole moiety's ability to mimic adenosine triphosphate (ATP), which is the natural substrate for kinases to form hydrogen bonds in kinase hinge regions. Docking was conducted using both rigid and flexible receptor conditions. Binding pose accuracy was quantified using atom-to-atom root mean square deviation (RMSD) and center-of-mass (CoM) RMSD, with 2 Å set as the success threshold. GNINA 1.0, which incorporates a 3D convolutional neural network (CNN)-based scoring function, achieved the highest atom-to-atom RMSD success rate (85.29%) under rigid docking. DOCK 6 performed second-best by exhibiting 79.71% success in rigid docking, but reduced to 61.19% while using a flexible docking approach. AutoDock Vina demonstrated comparable performance in both docking modes, with a success rate of 62.69% under rigid docking and 60.66% under flexible docking, indicating minimal variation between the two modes compared to the broader differences observed with other docking tools. AutoDock4 demonstrated moderate performance (<50% success) in both rigid and flexible modes. The accuracy of pose prediction varied by inhibitor class. Type 1 and Type 3 inhibitors were predicted with higher fidelity compared to Type 2 inhibitors, due to differences in binding site rigidity, hydrophobic interactions, and DFG-loop dynamics. Redocking analyses also assessed whether docking tools could recover three key interaction features: hinge hydrogen bonding, hydrophobic contacts, and DFG-loop orientations. GNINA 1.0 consistently performed well in recovering these features with lower computational cost. Interestingly, Vina showed better performance in terms of CoM RMSD but lower and consistent for atom-to-atom RMSD. Our study underscores the importance of scoring functions, receptor flexibility, RMSD type, and inhibitor class in determining docking accuracy for kinase-targeted drug design.