Protein trajectories get an order, not just snapshots: two-scale autoregression cuts 400-frame drift from 8.62 to 1.78 Å

Protein conformations are no longer delivered only as static structures or conformational ensembles; a backbone trajectory can be generated frame by frame in temporal order and used to compare conformational pathways across proteins, mutants, or ligand states. Earlier deep generative methods such as AlphaFlow and BioEmu emphasized sampling diverse conformational ensembles, and MDGen and similar work only began modeling sequential trajectories; ensemble generation reproduces a distribution of structures, whereas trajectory generation must also preserve temporal dependence between frames. On held-out ATLAS proteins, Pearson correlations were 0.91 for ensemble-averaged pairwise RMSD and 0.89 for residue-level RMSF; the leading principal-component direction agreed with MD in 76% of cases, versus 44% for AlphaFlow.

Error accumulation over long rollouts is split between two temporal resolutions: the slow scale carries long-range evolution, and the fast scale resolves structural variation between anchors. A single-scale model must balance short-range fluctuations against stable long-range evolution in every update, and as rollouts lengthen each prediction enters the next context, so geometric error accumulates. Against a single-scale model rolled out for 400 frames, the multiscale model reduced long-horizon RMSD error from 8.62 to 1.78 Å, raised pairwise-RMSD correlation from 0.14 to 0.77, and raised global RMSF correlation from 0.15 to 0.67; a 400-frame rollout took under 1 min, about 750-fold less than AlphaFlow under the reported evaluation settings.

Generation can be steered at inference: given a target conformation, the trajectory is biased toward that endpoint, and local DFG motif backbone geometry can be pushed toward a specified criterion without updating model parameters. Earlier conformational generators output samples from a distribution and offer no explicit control over the endpoint or over local motif geometry. Ten 500-frame rollouts from closed 1AKE conditioned on open 4AKE progressed along a shared tIC1 coordinate with frame order, with closed-state-specific contacts decreasing and open-state-specific contacts accumulating; for 68 of 100 kinases, DFG-in-guided and DFG-out-guided rollouts each yielded at least one accepted frame.

The demonstrated scope is backbone-centric: the model omits side chains, ligands, membranes, ions, and solvent. The 1-ns and 20-ns frame labels reflect the sampling strides used in training rather than calibrated kinetic time, and conditioned or steered rollouts are not unbiased samples. The intended readers are researchers studying protein conformational dynamics and kinase regulation, and the setting is rapid comparison of ordered conformational changes across proteins, mutants, or ligand states. The authors propose extending the model to include environmental context and validating generated states with state-discriminating observables, energetic calculations, or experiments as important next steps.

The authors state that the model remains backbone-centric, that frame labels are not calibrated kinetic time, and that conditioned and steered rollouts are not unbiased samples. The DFG analysis evaluates controllability rather than spontaneous switching, and 32 kinases did not yield both frame classes under the current rollout length and steering protocol. Kinase-panel Q-scores are treated as continuous similarity rather than state assignments because active and inactive structures share a large contact core; the crystallographic B-factor comparison reflects relative spatial patterns rather than absolute solution-state fluctuations. The GPCR analysis does not evaluate membrane-, ligand-, solvent-, or ion-dependent dynamics.