TEMPO

TEMPO is a generative model for protein conformational ensembles: instead of predicting a single static structure, it samples the range of shapes a protein adopts as it moves. Protein function is often governed by dynamics, but generating realistic ensembles is hard because molecular motion spans a wide range of timescales, from fast local fluctuations to slow, large-scale collective transitions. Physics-based molecular dynamics simulation captures this but is computationally expensive, motivating fast learned surrogates.

TEMPO frames protein dynamics as a Markovian process and generates trajectories autoregressively across two scales. A low-resolution model captures the slow, collective motions that move large parts of the structure, while a high-resolution model fills in detailed local fluctuations, yielding a temporal multi-scale factorization of the conformational landscape. Because it learns the dynamics rather than memorizing specific structures, TEMPO generalizes to proteins it has never seen during training.

TEMPO was introduced in late 2025 by Yaoyao Xu, Di Wang, Zihan Zhou, Tianshu Yu, and Mingchen Chen, a collaboration between the School of Data Science at The Chinese University of Hong Kong, Shenzhen, and Changping Laboratory in Beijing. It was presented as a NeurIPS 2025 poster and is described in an accompanying arXiv preprint.

#Key Features

• Conformational ensemble generation: TEMPO samples ensembles of structures rather than a single conformation, targeting the dynamic behavior that underlies much of protein function.
• Temporal multi-scale factorization: A low-resolution model captures slow collective motions while a high-resolution model generates fine local fluctuations, separating the timescales of protein motion.
• Markovian autoregressive generation: Dynamics are modeled as a Markov process and trajectories are generated step by step, so each frame is conditioned on the previous state.
• Zero-shot transfer: Trained on molecular dynamics trajectories with low sequence overlap to the test sets, TEMPO generalizes to 64 held-out proteins without retraining.
• Fast surrogate for MD: The model generates a 400-frame trajectory in roughly 0.006 hours, providing a rapid alternative to explicit molecular dynamics simulation.

#Technical Details

TEMPO is a hierarchical autoregressive framework combining multi-head attention, GRU layers, and Invariant Point Attention (IPA) modules to generate conformations sequentially under a Markov assumption. It is trained on molecular dynamics trajectory data: from mdCATH, 1,000 proteins (sequences truncated to 240 residues, trajectories standardized to 400 frames at 1 ns intervals), and from ATLAS following the protocol of the MDGen baseline, with 64 proteins held out for testing on mdCATH and 50 reserved for validation. Sequence similarity between train and test sets is kept well below standard thresholds (18.3% for ATLAS), so evaluation reflects generalization to unfamiliar proteins. On mdCATH, TEMPO reaches a pairwise-RMSD correlation of 0.77 with an RMSD error of 1.78 Angstrom; on ATLAS it reaches a pairwise-RMSD correlation of 0.91 with an RMSD error of 1.83 Angstrom, while generating a full 400-frame trajectory in about 0.006 hours.

#Applications

TEMPO is aimed at structural biologists and computational researchers who need to explore protein flexibility, such as sampling alternative conformations for allostery, cryptic-pocket discovery, or ensemble-based docking, without running costly molecular dynamics campaigns. Because it transfers zero-shot to new sequences, it can be applied to proteins for which no trajectory data exist, providing a fast first look at likely conformational behavior that can then be refined with targeted simulation.

#Impact

TEMPO contributes to a growing class of learned surrogates for molecular dynamics that aim to make conformational ensemble generation tractable at scale. Its explicit separation of slow collective motions from fast local fluctuations offers a principled way to cover multiple timescales in a single generator, and its zero-shot performance on held-out proteins indicates the approach captures transferable features of protein dynamics. As a recent conference preprint, its results are computational and await broader peer review and experimental comparison; the paper does not report an associated public code release.