Year: 2026
Type: preprint
Abstract: Protein generators increasingly produce conformational ensembles and trajectories as faster alternatives to molecular dynamics (MD). Yet agreement with an equilibrium ensemble does not reveal how conformations interconvert. We show that ensemble evidence can survive even when temporal dynamics are destroyed. Randomly reordering authentic MD frames leaves ensemble fidelity unchanged but reduces the post-gate kinetic pass rate from 0.963 to 0.004; an independent MD replicate reaches 0.95 under the same test. Dynbench separates conformational coverage, time-dependent behaviour and geometric and fluctuation 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-centred evaluation therefore systematically overstates evidence of learned dynamics. Claims about protein dynamics require direct temporal validation at a resolvable physical interval.
Authors: Kaining Liu, Qiuting Qian, Ying Chi
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7208830581