Molecular docking is a cornerstone of structure-based drug design but often fails to reliably identify the correct binding orientation of ligands, as commonly used scoring functions neglect protein flexibility and dynamic solvation effects. It is commonly assumed that a treatment of protein flexibility, e.g. by running molecular dynamics simulations, offers a more robust physics model of protein-ligand complexes. Here, we evaluate whether molecular dynamics (MD) simulations, run as a post-docking validation method, actually can confirm correct poses and flag incorrect ones. A curated benchmark of several hundred proteinligand complexes was assembled and automatically prepared for modelling. A semiautomated docking workflow generated both near-native and decoy poses to subject to explicit-solvent MD simulations. Pose stability and quality after MD were then evaluated. The findings confirm that short, GPU-accelerated MD simulations provide a simple and computationally feasible refinement step after docking.