ESMDynamic

Structure prediction gave biology a static picture of the proteome, but function lives in motion: gates that open, loops that swing, interfaces that form and break. Generative ensemble models can sample that motion, yet they pay for it with hundreds of forward passes per protein. ESMDynamic, from Diwakar Shukla's group at the University of Illinois Urbana-Champaign, takes a different target. Instead of generating conformations, it predicts the dynamic contact map directly — for every residue pair, the probability that the pair fluctuates between contact and non-contact, plus contact frequencies reflecting equilibrium populations and coarse-grained formation and dissociation kinetics.

The model is built on ESMFold, whose weights stay frozen. What is trained is a small head that consumes ESMFold's auxiliary outputs — distogram logits, language model logits, pTM and pLDDT — biases the sequence and pair representations from the final Evoformer block, and passes them through a two-block Dynamic Contact Module. Because the backbone needs no multiple sequence alignment, a prediction costs a single forward pass from a single sequence.

Posted as a preprint in August 2025, ESMDynamic has been applied zero-shot to systems well outside its training distribution and run across more than 18,000 human proteins, positioning it as a general sequence-to-dynamics annotator rather than a per-target predictor.

#Key Features

• Dynamic contacts, not conformations: Output is a probabilistic map over residue pairs distinguishing fluctuating contacts from static ones and from non-interacting pairs, along with contact frequencies and coarse-grained kinetics across temperature conditions.
• Accuracy against MD ground truth: On a sequence-identity-aware test split with under 20% identity between train, validation, and test clusters, it reaches 80% balanced accuracy and 73% recall, with over 90% of positive predictions falling within five residues of a true positive.
• Orders-of-magnitude speedup: Predicting the 82-protein ATLAS test set took 9.45 minutes on a single NVIDIA RTX 4090; the fastest generative baseline needed roughly 7.9 hours at 250 samples per protein. Runtime is essentially that of ESMFold itself.
• Generalization beyond the training set: Membrane transporters, a de novo designed protein, and a homodimeric complex are all handled zero-shot, despite the training data being biased toward ordered, soluble, monomeric domains.
• Collective variable selection: Predicted dynamic contacts can be clustered to nominate residue-residue distances as collective variables for Markov state model construction, automating a step that normally depends on expert intuition.

#Technical Details

Training proceeded in two stages, both on a single RTX 4090. Pretraining used dynamic contacts inferred from structural variation within 88,294 sequence-aligned clusters (above 95% intracluster identity) drawn from experimentally resolved structures and AlphaFold DB models — 946,397 chains from 598,144 unique structures, a median of four conformers per protein. The model was then fine-tuned on mdCATH, 5,398 soluble domains with a mean trajectory length of 464 ns, using only the 320 K simulations. Both stages crop proteins to 256 residues, use Adam at a learning rate of 1e-4, and optimize a focal loss (γ = 2; α = 0.99 for pretraining, 0.85 for fine-tuning) to counter severe class imbalance. On the mdCATH and ATLAS test sets, ESMDynamic matches or exceeds AlphaFlow, ESMFlow, and BioEmu at identifying dynamic contacts — notable on ATLAS, where AlphaFlow trained on that data and ESMDynamic had never seen it.

#Applications

The model fits naturally upstream of simulation. Predicted dynamic contacts can seed enhanced-sampling collective variables, focus MD analysis on the residue pairs that actually move, and triage which targets in a set warrant expensive trajectory generation at all. For protein engineering, the map flags flexible loops and gating elements to stabilize or exploit; the paper demonstrates this on the ASCT2 and SWEET2b transporters, a de novo troponin C design, and the HIV-1 protease homodimer. At proteome scale, the released human predictions covering more than 18,000 proteins support comparative analysis of conformational variability across families. A Colab notebook and a Docker image with automatic weight download make the model runnable without local environment work.

#Impact

ESMDynamic reframes protein dynamics prediction as a discriminative annotation problem rather than a generative sampling one, and the trade is favorable when the question is which contacts move rather than what does the ensemble look like. That it was trained on a single consumer GPU on top of a frozen backbone is itself a useful result about how cheaply dynamics signal can be extracted from an existing structure predictor. Code is released under the MIT license with a downloadable pretrained checkpoint, and the featurized training datasets and weights are deposited in the Illinois Data Bank. The limitations are stated plainly by the authors: training data skew toward ordered, monomeric, well-sampled proteins, multimeric systems are scarce, and the model reports contact-level dynamics rather than the full conformational ensemble, so it complements rather than replaces simulation and generative ensemble methods.