Which Metrics Best Capture Protein Structural Changes in Molecular Dynamics Simulations? Evaluating Score Combinations and Force-Field Effects

The scores RMSD, TM-score, GDT, lDDT, SphereGrinder, CAD, QCS, FlexE, and MolProbity provide an automated, comprehensive assessment of conformational changes and structural quality of proteins and thus represent a promising tool for routine use in molecular dynamics (MD) simulations, particularly in high-throughput settings. However, it remains unclear to what extent these scores provide redundant information and which scores are most informative for capturing conformational changes. Based on MD simulations of 268 diverse proteins, we demonstrate that the investigated scores are highly correlated and that one global score (or QCS), one local score, and FlexE capture almost 90% of the variance across all scores. Since MD randomness partially explains FlexE’s variability, we argue that a combination of one global and one local score is sufficient for most practical applications. We also investigate the influence of simulation setups and find that the choice of the force field can significantly affect scoring results. In particular, the setup using the Amber ff19sb force field yields systematically different scores than all CHARMM36m-based setups, emphasizing the importance of methodological choices when designing MD experiments and interpreting their results.