Evaluating Performance of Rigid and Flexible Docking Strategies for BACE1 Complexes: A Comparative Study of Physics-based and AI-Driven Docking Methods

Beta-site amyloid precursor protein cleaving (BACE1) enzyme 1 is a critical therapeutic target in Alzheimer’s disease, yet its dynamic active site poses significant challenges for accurate binding pose prediction for small molecules. This study benchmarked physics-based, deep learning-based, and generative molecular docking tools for binding pose prediction using both rigid and flexible receptor protocols for ~431 BACE1 ligand complex structures from the Protein Data Bank. Docking tools were compared based on all-atoms Root Mean Square Deviation (RMSD) and centre-of-mass RMSD deviation between the experimental and predicted binding poses. Physicochemical parameters, including ligand flexibility, solvent-accessible surface area, ligand polarity, and protein hydrophobicity, were evaluated to determine their influence on RMSD and potential bias in tool performance. DOCK6 demonstrated the most reliable performance, mainly due to its grid-based scoring and anchor-and-grow sampling strategy, though it lacks flexible receptor docking. GNINA, a deep learning-based scoring method, is known to be highly effective in binding pose prediction, showed reduced accuracy, likely due to under-representation of this target in its training data. Flexible docking protocols struggled with pose ranking and geometry, particularly for large or flexible ligands, while generative models like DiffDock often failed to produce native-like poses. This study provides the first systematic benchmarking framework for BACE1 docking, offering practical guidance for method selection and a basis for improving flexible and AI-driven docking strategies in structure-based drug design. Given the critical need for effective BACE1 inhibitors after multiple clinical failures, these insights could accelerate next-generation therapeutic discovery with enhanced efficacy and reduced off-target effects.