Ensemble tests mask missing dynamics in protein conformational generators

Abstract

Protein generators are increasingly used to produce conformational ensembles and trajectories as faster alternatives to molecular dynamics (MD). Yet whether agreement with an equilibrium ensemble demonstrates that a generator has learned dynamics remains unresolved. Here we show that this evidence survives the destruction of dynamics. Randomly reordering authentic MD frames leaves ensemble fidelity unchanged but reduces the geometrically gated kinetic pass rate from 0.963 to 0.004; an independent MD replicate reaches 0.95 under the same test. Dynbench separates recovery of conformational states, time-dependent behaviour and trajectory admissibility. Public generators that appear similar under ensemble tests separate sharply, and in a prospective 36-protein lockbox the reference and floor remain stable while two models reverse order. Ensemble tests can therefore systematically overstate evidence of learned dynamics, establishing time-resolved validation as a distinct requirement for claims about protein dynamics generators.