DynaFold

DynaFold is a generative framework that samples all-atom protein dynamics — trajectories, conformational ensembles, and transition pathways between states — using latent diffusion. It was developed by Zirui Fan, Junjie Zhu, and Hai-Feng Chen at Shanghai Jiao Tong University and posted to bioRxiv in September 2025.

The problem it targets is the cost of molecular dynamics (MD). MD produces time-resolved, Boltzmann-distributed trajectories at atomic resolution, but all-atom simulation scales poorly with system size, and coarse-grained force fields buy speed by discarding atomic detail. A generation of deep learning emulators has attacked this from the ensemble side — AlphaFlow fine-tunes a structure predictor under a flow matching objective, and BioEmu-1 emulates equilibrium distributions directly — but these sample independent structures rather than continuous trajectories, and none reproduce the pathway a protein takes between two conformations.

DynaFold's design choice is to move the diffusion process off the 3D coordinates entirely. A variational autoencoder first compresses backbone structure into a low-dimensional continuous latent space; a Latent Denoising Transformer (LDT) then runs diffusion inside that space using ordinary self-attention, with no equivariant or geometric machinery. Because the VAE maps chemically related conformations to nearby latent points, the model can interpolate between states without producing the physically impossible geometries that direct interpolation in Cartesian space yields. The framework also drops the MSA dependency common to structure predictors, accepting either an initial backbone or a single sequence embedded by a protein language model.

#Key Features

• Diffusion in a learned latent space: A VAE compresses each residue to a 12-dimensional continuous representation, letting the denoiser operate on a plain 2D matrix with standard attention instead of SE(3)-equivariant layers.
• Three task modes: Separate checkpoints handle forward simulation (initial structure to trajectory), conformational transition (sampling paths between two given endpoints), and sequence-to-ensemble prediction.
• Continuous transition pathways: Unlike ensemble samplers, DynaFold generates ordered, time-resolved paths through folding, unfolding, and α-to-β transitions, and can sample multiple distinct routes between the same two states.
• MSA-free inputs: Conditioning uses backbone coordinates or single-sequence embeddings, removing the alignment search that dominates runtime in MSA-based predictors.
• All-atom output: The decoder reconstructs full side-chain coordinates rather than backbone-only frames.

#Technical Details

The VAE encoder adapts the VQ-VAE architecture from ESM-3, replacing the discrete codebook with continuous mean and variance outputs (31M parameters); a stacked-Transformer decoder predicts Cartesian atom coordinates (158M parameters). The LDT combines an Input Embedder with Dynaformer blocks that alternate spatial and temporal attention, trained under the EDM diffusion framework. The forward-simulation and conformational-transition LDTs use 26M parameters (hidden dimension 384); the ensemble model scales to 306M (hidden dimension 1024) and takes ESM-2 650M token embeddings as sequence input. Training used 4 NVIDIA A100 40GB GPUs.

The VAE was trained on 231,751 PDB chains plus 704,391 AlphaFold Database structures filtered by pLDDT and long-range contact count; the LDT was trained on MD trajectories from ATLAS (1,267 training, 69 held-out test) and the D. E. Shaw Research fast-folding set (11 proteins, evaluated by leave-one-out cross-validation). On the 69 ATLAS test proteins the preprint reports lower maximum mean discrepancy to MD distributions than AlphaFlow and BioEmu-1, with roughly a tenfold reduction in sampling time relative to AlphaFlow, and larger RMSF and pairwise RMSD than the comparably sized MDGen trajectory model. Across fast-folding proteins, sampled conformations reached MD free-energy minima at an average backbone RMSD of 2.68 Å.

#Applications

The model suits researchers who need conformational diversity faster than MD can supply it: identifying cryptic or transient pockets for structure-based drug design, characterizing allosteric states, and studying folding and unfolding mechanisms where the ordering of events matters. The transition-pathway mode addresses a question ensemble predictors cannot answer — given two known functional states, what does the route between them look like — which is relevant to conformational-change-driven targets such as transporters and kinases.

#Impact

DynaFold demonstrates that latent diffusion, well established in image generation, transfers to protein dynamics and that a 26M-parameter denoiser can outperform larger geometry-aware architectures once the representation is right. The authors are candid about the limits: free-energy surfaces remain inaccurate, β-sheet folding is reproduced poorly, side-chain χ2 through χ4 torsions are not captured, and the VAE trained on crystal structures cannot model side-chain dynamics, so all-atom MD remains preferable where those details matter. The work is a preprint awaiting peer review. Source code and model weights are released under the MIT license.