DynaProt

Molecular dynamics simulation is the in-silico reference standard for studying how proteins move, but the cost of running trajectories long enough to characterize an ensemble puts it out of reach at proteome scale. The deep-learning response has largely been generative: models such as AlphaFlow and BioEmu-1 sample many plausible conformations and let the spread of those samples stand in for dynamics — which means paying for a large network and hundreds of forward passes per protein before any dynamics quantity can be measured.

DynaProt takes the opposite route. Developed by Mihir Bafna, Bowen Jing, and Bonnie Berger at MIT CSAIL and posted as a preprint in September 2025, it treats a conformational ensemble as a multivariate Gaussian over Cα coordinates and predicts that distribution's parameters directly from a single static structure — no sampling loop, no per-protein refitting. The paper was accepted to ICLR 2026.

Dynamics are estimated at two complementary scales. Per-residue 3×3 marginal covariance matrices describe local flexibility as anisotropic "blobs," capturing the direction of motion and not only its magnitude; RMSF falls out as the square root of each matrix's trace. A separate N×N scalar coupling matrix encodes how each residue's motion correlates with every other residue's. Composed, the two yield an approximate 3N×3N joint covariance that can be sampled to produce ensembles in a fraction of a second.

#Key Features

• Gaussian parameterization of the ensemble: Rather than generating structures and measuring their variance, the model predicts the covariance itself, turning ensemble characterization into a single deterministic forward pass.
• Anisotropic per-residue flexibility: The 3×3 marginals encode directionality, so a hinge that swings along one axis is distinguished from a loop that fluctuates isotropically — information a scalar RMSF value discards.
• Residue–residue dynamic coupling: The pairwise head predicts correlated motion across the chain, the signal underlying allosteric communication and concerted domain movement.
• Sub-second ensembles: Sampling takes roughly 0.14 seconds per protein, over 70,000× faster than the template-conditioned AlphaFlow-MD baseline.
• Small enough to be cheap: At 2.86 M parameters total, DynaProt is orders of magnitude smaller than the 1.4 B-parameter generative predictors it outperforms on RMSF correlation.
• SE(3)-invariant by construction: Predictions are made in local residue frames, so global rotations and translations of the input leave the learned representations unchanged.

#Technical Details

Both heads share a backbone of eight Invariant Point Attention blocks taken from the AlphaFold 2 structure module, operating on local Cα residue frames. The marginal module (DynaProt-M, 955 K parameters) is an MLP readout predicting six Cholesky factors per residue, guaranteeing a positive semi-definite 3×3 covariance. The joint module (DynaProt-J, 1.9 M parameters) projects concatenated residue pairs into a stack of Evoformer-style pairwise attention blocks with triangle updates before a Cholesky reconstruction of the coupling matrix; OpenFold components are vendored under Apache 2.0.

Training uses the ATLAS molecular dynamics dataset — 1,390 proteins chosen for structural diversity by ECOD domain classification, each with three 100 ns replicate simulations — under AlphaFlow's 1265/39/82 split, with no large-scale PDB pretraining. On the ATLAS test set, DynaProt reaches a global RMSF Pearson r of 0.71 and a per-target r of 0.86, ahead of ConfDiff (0.67 / 0.85), AlphaFlow-MD (0.63 / 0.77), and BioEmu (0.19 / 0.67). On the FlexPert evaluation split, median RMSF Pearson r is 0.865 against 0.830 for the 1.2 B-parameter FlexPert-3D and 0.697 for anisotropic network model normal mode analysis. The trade-off is distributional: on metrics probing the shape of the landscape — Wasserstein distance in MD principal component space, transient contact recovery — the much larger AlphaFlow-MD baseline remains ahead, since one Gaussian cannot represent discrete open and closed states.

#Applications

The speed makes flexibility a property that can be annotated in bulk — across the PDB, an AlphaFold DB proteome, or a design campaign's output — rather than measured one protein at a time. The command-line tool writes per-residue RMSF as a CSV and into a PDB B-factor column, exports ensembles as PDB or XTC, and renders the marginals as PyMOL ellipsoids. Uses include triaging which regions warrant expensive MD, flagging flexible loops and hinges before crystallization or cryo-EM, reading the coupling matrix for candidate allosteric paths, and guiding rigidification for thermostability engineering.

#Impact

DynaProt reframes ensemble prediction as direct regression on ensemble statistics rather than a generative sampling problem, showing that a 2.86 M-parameter model trained only on ATLAS can match or beat billion-parameter samplers on residue-level flexibility at a tiny fraction of the compute. Code is released under the MIT license with checkpoints on HuggingFace, lowering the barrier to using it as a cheap dynamics prior inside larger pipelines. The limitations are inherent to the formulation: predictions are Cα-only and single-chain, the Gaussian assumption cannot capture multimodal conformational change, and the training signal comes from 100 ns simulations, so slower transitions are outside what the model has learned to describe.