Dyna-1

Proteins are not static structures. Many function through motions on the microsecond-to-millisecond (µs-ms) timescale, in which parts of the chain interconvert between conformational states. This "conformational exchange" underlies enzyme catalysis, allostery, and ligand binding, but it is largely invisible to structure prediction tools that return a single folded model. Measuring µs-ms dynamics experimentally, typically with NMR relaxation-dispersion methods, is slow and demands specialist expertise, so dynamics data exists for only a small fraction of characterized proteins.

Dyna-1 addresses this gap by predicting, for each residue in a protein, the probability that it experiences µs-ms motion, taking sequence and/or structure as input. The model comes from the Wayment-Steele Lab at Scripps Research and was introduced in the 2025 preprint "Learning millisecond protein dynamics from what is missing in NMR spectra." Its central insight is that residues which go missing, that is, unassigned, in NMR spectra are frequently exchange-broadened by µs-ms motion. Rather than treating those absent peaks as a nuisance, the authors use them as a training signal, turning a large body of existing NMR data into supervision for a dynamics predictor.

Dyna-1 is a frozen-checkpoint predictor: pretrained weights are downloaded and applied zero-shot to arbitrary new proteins through a standalone inference script, with no per-protein retraining.

#Key Features

• Per-residue exchange probability: For any input protein, Dyna-1 returns a per-residue probability that the position undergoes µs-ms conformational exchange, providing a residue-level map of predicted dynamics.
• Sequence and/or structure input: Predictions can be driven from sequence alone or from sequence together with structure, so the model applies to proteins with or without an experimental or predicted structure.
• Learning from missing NMR peaks: The training signal exploits residues that lack NMR assignments, which are enriched for exchange broadening, converting incomplete spectra into usable labels.
• Protein language model backbone: Dyna-1 builds on an intermediate representation (layer 22) of the ESM-3 esm3-sm-open-v1 model, with an alternate variant built on ESM-2 embeddings.
• Open code and downloadable weights: The GitHub repository provides an MIT- licensed inference workflow and pretrained weights; because the primary model depends on ESM-3 weights, those weights are distributed under EvolutionaryScale's non-commercial license.

#Technical Details

Dyna-1 attaches a prediction head to features drawn from an intermediate transformer layer of a pretrained protein language model, ESM-3 (esm3-sm-open-v1, layer 22), with an ESM-2-based alternate provided for users who need it. Training uses two curated resources released alongside the model: RelaxDB, a collection of 133 NMR relaxation datasets (R1, R2, and heteronuclear NOE measurements) assembled from the BMRB and the literature, and RelaxDB-CPMG, 10 CPMG relaxation-dispersion datasets. Missing backbone amide assignments serve as an additional dynamics label. Across more than 100 NMR datasets analyzed in the study, the model's predictions correlate with experimentally measured exchange, with the signal most pronounced at residues implicated in enzyme catalysis and ligand binding. Weights and the RelaxDB datasets are hosted on HuggingFace, and inference runs from the command line on a supplied PDB and chain.

#Applications

Dyna-1 gives biochemists and structural biologists a fast, sequence-level screen for where a protein is likely to be dynamic, complementing static predictors such as AlphaFold that describe a single conformation. Predicted high-exchange residues can prioritize sites for mutagenesis, help interpret enzyme mechanism and allostery, flag candidate cryptic pockets or binding-coupled regions, and guide the design of NMR relaxation experiments toward the proteins and residues most likely to reveal functionally important motion.

#Impact

By reframing the peaks that are absent from NMR spectra as a rich source of dynamics information, Dyna-1 makes a previously data-starved property, µs-ms conformational exchange, predictable at proteome scale from sequence. The finding that predicted dynamics concentrate at catalytic and binding residues, and appear evolutionarily conserved, reinforces the view that motion is an intrinsic part of protein function rather than incidental flexibility. As a preprint, the work awaits peer review; the model reports the probability of µs-ms exchange rather than full exchange kinetics or thermodynamics, and its primary weights inherit the non-commercial terms of the ESM-3 backbone.