An AI model built to simulate how proteins move has learned to stop retracing its own steps.
Generative emulators of protein dynamics are fast stand-ins for molecular dynamics simulations, but they inherit the limits of their training data. Left running long enough, they tend to loop back through states they have already generated rather than discover new ones. Researchers added a history-aware bias to a frozen, pretrained emulator that nudges its sampling away from previously seen structures, echoing a trick long used in classical enhanced-sampling methods. A separate refinement step then snaps any drifted structures back onto biologically valid shapes. On the DynamicPDB-80 benchmark, the method boosted structural diversity by 35%, and on 12 proteins the model had never seen during training, it reached the same coverage of protein states up to 37 times faster while finding roughly three times as many low-energy, and often more biologically relevant, conformations.
The result matters because protein simulation tools usually force a choice between speed and completeness. Fast generative models miss rare folding intermediates and binding pockets that slower, more rigorous simulations can find, which is a real problem for drug discovery pipelines that depend on those rare states. Bolting a bias onto an already-trained model, instead of retraining it from scratch, is a cheap way to close that gap.
The authors say code is coming "soon," arXiv's classic placeholder for "not yet," so independent verification will have to wait.