Abstract
Conformational sampling remains a major challenge in molecular dynamics (MD) simulations of proteins with large-scale functional motions. Here, we use apo Escherichia coli adenylate kinase (AdK) as a benchmark for the open–closed transition and compare nine sampling protocols under a uniform explicit-solvent budget of 300 ns per method. The methods include unbiased MD, metadynamics, targeted MD (TMD), accelerated MD, simulated tempering, interpolation-based MD, and BioEmu. Closed-state sampling was evaluated relative to a reference-closed basin defined from ligand-stripped PDB 1AKE. TMD was the only method that both reached and maintained the closed reference, whereas unbiased MD did not enter this basin within 300 ns. Metadynamics increased sampling along the domain-distance coordinate but also promoted excessive LID opening, while simulated tempering showed a similar tendency. BioEmu generated closed-like conformations, but these states were not maintained during subsequent short MD simulations. Overall, the results show that collective-variable coverage does not necessarily correspond to stable target state occupancy and that different sampling methods provide different types of conformational information. These findings highlight the importance of selecting sampling strategies according to the specific scientific objective.