ProTDyn

Molecular dynamics (MD) simulation remains the reference method for exploring how proteins move between conformational states, but its integration timestep must be orders of magnitude smaller than the events of interest, making long simulations prohibitively expensive. Deep generative models have attacked this bottleneck from two directions that have stayed separate: learning the equilibrium Boltzmann distribution over conformations, and learning the transition density between successive frames. ProTDyn, developed at Purdue University with a collaborator at Northwestern University, unifies both inside a single pretrained autoregressive protein language model.

The key move is representational. ProTDyn encodes a conformation as a sequence of discrete structure tokens, which turns ensemble sampling and trajectory propagation into the same next-token prediction problem, differing only in what the model conditions on. Three sampling modes fall out of that formulation: independent and identically distributed sampling of equilibrium structures, forward trajectory generation at a chosen timestep, and "dynamics inpainting," which recovers fine-grained frames inside a coarsely sampled trajectory. Because training spans trajectories at 1 ns, 10 ns, and 100 ns resolution, one checkpoint covers transitions from nanoseconds to microseconds.

Released as a preprint in September 2025 and accepted to ICLR 2026 as a main conference paper, ProTDyn benchmarks itself directly against BioEmu, the closest equilibrium-ensemble emulator, and adds the temporal dimension that ensemble-only generators such as AlphaFlow and Distributional Graphormer do not model.

#Key Features

• Unified thermodynamics and dynamics: One set of weights samples equilibrium ensembles consistent with Boltzmann statistics and propagates time-correlated trajectories, rather than requiring separate specialized models.
• Multi-timescale generation: The dynamics module is trained at 1 ns, 10 ns, and 100 ns timesteps with a 10-frame memory kernel, giving effective horizons of 10 ns, 100 ns, and 1 µs from the same checkpoint.
• Dynamics inpainting: Coarse 100 ns strides are refined into ten 10 ns sub-intervals. This combination is more accurate than stepping at 10 ns directly, because long strides require fewer autoregressive steps and accumulate less error.
• Exact likelihoods: Autoregressive modeling over discrete tokens yields exact likelihoods for ensembles and trajectories — a quantity diffusion- and flow-based generators can only approximate, often at prohibitive cost.
• Transfer to unseen proteins: A fixed pretrained checkpoint is applied to proteins outside the training set with no per-system refitting.

#Technical Details

Conformations are discretized with the pretrained ESM-3 structure tokenizer, which assigns each residue one of 4,096 structure tokens via a VQ-VAE encoder, with a paired decoder reconstructing 3D coordinates. The ESM-3 sequence and structure embedding modules are frozen throughout training. The backbone is 24 transformer blocks totaling 1.4 billion parameters, following ESM-3 in using Pre-LN, SwiGLU activations, and rotary embeddings, and extending them with a two-layer rotary scheme in which one layer encodes residue position along the chain and the other encodes temporal position at a smallest unit of 1 ns.

The training corpus combines 542,378 sequence–structure pairs from the Swiss-Prot subset of the AlphaFold Database with equilibrium MD from mdCATH (5,398 proteins simulated up to 500 ns) and the BioEmu corpus, which contributes 1,100 octapeptides at 5 µs each, 50 CATH domains at 100 µs, 1,100 CATH domains at 39 µs, and MEGAsim trajectories. On the 50-domain CATH benchmark, which the paper notes is included in the thermodynamics training set and serves primarily as a comparison against the baseline, the thermodynamics module reaches Jensen–Shannon divergences of 0.023 for radius of gyration, 0.012 for RMSD to the AlphaFold 3 native structure, and 0.155 across the top two TICA components, against 0.082, 0.137, and 0.293 for BioEmu. In Markov state model comparisons, 50 µs of generated 100 ns-step trajectory recovers stationary distributions and transition probabilities at roughly the accuracy of 50 µs of reference MD.

#Applications

Researchers who need conformational ensembles rather than a single static structure are the primary beneficiaries. In structure-based drug discovery, generated ensembles expose cryptic binding pockets that open only transiently, support induced-fit and ensemble-docking workflows, and allow candidates to be scored against realistic conformational populations. The dynamics modules extend this to questions requiring time ordering: allosteric communication, transition pathways between metastable states, and residence times implied by a Markov state model. Enzyme engineering and intrinsically disordered proteins, where a single predicted fold is uninformative, are further targets.

#Impact

ProTDyn demonstrates that discrete structure tokenization plus autoregressive language modeling is a workable substrate for protein dynamics, not just static structure, and that thermodynamic and dynamic training reinforce each other within one architecture. Code is MIT-licensed and the 16 GB checkpoint is publicly downloadable from Hugging Face. Limitations are acknowledged by the authors: performance is bounded by the scarcity of long equilibrium MD data, the memory kernel at each timescale is set by hand rather than learned, and generation at short timesteps accumulates autoregressive error that damps long-timescale correlations. Validation covers CATH domains, octapeptides, and small proteins rather than large complexes, and is entirely in silico, benchmarked against reference simulations rather than experimental observables.