When does molecular dynamics improve RNA models? Insights from CASP15 and practical guidelines

Molecular dynamics (MD) simulations are increasingly applied to refine biomolecular models, yet their practical value in RNA structure prediction remains unclear. Here, we systematically benchmarked the effect of MD on RNA models submitted to the CASP15 community experiment, using Amber with the RNA-specific χ_OL3 force field. Across 61 models representing diverse targets, we find that short simulations (10-50 ns) can provide modest improvements for high-quality starting models, particularly by stabilizing stacking and non-canonical base pairs. In contrast, poorly predicted models rarely benefit and often deteriorate, regardless of their CASP difficulty class. Longer simulations (>50 ns) typically induced structural drift and reduced fidelity. Based on these findings, we provide practical guidelines for selecting suitable input models, defining optimal simulation lengths, and diagnosing early whether refinement is viable. Overall, MD works best for fine-tuning reliable RNA models and for quickly testing their stability, not as a universal corrective method.