Convergence is not correctness: context-dependent performance of enhanced-sampling methods across biological complexity

Enhanced-sampling molecular dynamics is indispensable for studying biomolecular processes beyond the reach of conventional simulation, yet its reliability in realistic biological settings remains poorly characterized. Here, we compare four widely used approaches—replica-exchange umbrella sampling (REUS), well-tempered metadynamics (WT-MtD), well-tempered metadynamics–extended adaptive-biasing force (WTM-eABF), and on-the-fly probability enhanced sampling (OPES)—across five experimentally grounded processes: ligand binding, conformational change, kinase activation, membrane permeation, and ion conduction. Three findings emerge. Method performance does not transfer across biological contexts; each approach exhibits characteristic failure modes invisible in low-dimensional model systems but consequential in realistic landscapes. Self-consistency metrics, the field’s standard convergence diagnostic, are insufficient: simulations can satisfy convergence criteria while yielding quantitatively incorrect free-energy profiles. Cross-method agreement against an independent reference is therefore necessary. These results re-frame method selection as a context-dependent decision and argue that cross-method validation should become a community standard. Enhanced-sampling molecular dynamics methods are widely used to probe biomolecular processes, yet their reliability in realistic biological systems remains poorly characterized. This study shows that method performance does not transfer across biological contexts, that standard convergence metrics can be misleading, and that reliable free-energy estimation requires cross-method validation, underscoring the need for context-aware method selection.