Use Dyna-1 Online

Dyna-1: Predicting Microsecond-to-Millisecond Protein Dynamics from Sequence and Structure

Dyna-1 is a deep learning model that predicts microsecond-to-millisecond (µs-ms) protein motions at the residue level. It was developed by Hannah K. Wayment-Steele, Gina El Nesr, Ramith Hettiarachchi, Hasindu Kariyawasam, Sergey Ovchinnikov and Dorothee Kern, with affiliations including Scripps Research and the Howard Hughes Medical Institute, Stanford University and MIT. The model builds on an intermediate layer of the multimodal language model ESM-3, and the code is released for noncommercial use.

Many proteins depend on switching between conformations on the µs-ms timescale for enzyme catalysis, ligand binding and allostery. Structure predictors such as AlphaFold 2 learned from the Protein Data Bank, which holds static structures without time information, and the authors note that dynamics prediction has lacked large standardized experimental benchmarks to train and test against.

The authors' key idea is that NMR experiments already contain a hidden signal. Residues whose motion broadens their NMR signals beyond detection are left unassigned in chemical shift datasets. By treating missing backbone amide assignments across roughly 10,000 deposited proteins as a marker of µs-ms exchange, they could train a model at a scale that relaxation experiments alone could not support.

Why Predicting Millisecond Protein Motion Matters

Motions on the µs-ms timescale are directly linked to biological function and are often concerted across a protein. In the authors' RelaxDB analysis, residues with µs-ms motion were more conserved than residues with no motion, which in turn were more conserved than residues with picosecond-nanosecond motion. That link between conservation and dynamics is what made an evolution-informed language model a natural fit for the task.

How Dyna-1 Learns Dynamics from What Is Missing in NMR Spectra

1. Curate relaxation data: the team compiled R1/R2/NOE datasets for 163 protein domains and labelled residues with exchange (Rex) by comparing experimental R2 with a rigid-tumbling prediction computed on AlphaFold 2 models.

2. Build a large training set: from the BMRB, they filtered out selectively labelled and deuterated entries and other outliers, leaving 9,381 proteins in which residues are labelled as assigned or missing.

3. Frame a classification task: a model predicts, for each residue, whether its backbone amide assignment is missing, used as a proxy for µs-ms exchange. Prolines and unassigned termini are excluded.

4. Compare architectures: AlphaFold 2 pair features, ESM-2 and ESM-3 embeddings were tested against a one-hot sequence baseline. Layer 22 of ESM-3, given sequence and structure, performed best and became Dyna-1.

Dyna-1 Performance and What It Predicts Well

The authors report that Dyna-1 also predicts µs-ms motion measured directly in NMR relaxation experiments, and that dynamics tied to function, including enzyme catalysis and ligand binding, are especially well predicted. Examples include the FMN-binding loops of flavodoxin YcqA and loops in M. tuberculosis tyrosine phosphatase A that differ between apo and holo structures. Revisiting published data also led to reinterpretations, such as proteins that bind phosphate yet were measured in phosphate buffer.

What is Tamarind Bio?

Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.

The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:

• A user-friendly graphical interface for setting up and launching experiments

• A robust API for integration into existing research pipelines

• An automated system for managing and scaling computational resources

Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.

Accelerating Discovery with Dyna-1 on Tamarind Bio

• Locating functional motion: flag loops and regions likely to undergo µs-ms exchange, such as ligand-binding sites and catalytic loops.

• Prioritizing NMR experiments: identify residues likely to be exchange-broadened before choosing relaxation or CPMG experiments.

• Interpreting existing NMR data: compare predicted exchange with measured relaxation data to spot residues worth re-examining.

• Linking dynamics to conservation: the authors found predictions are strongest for conserved residues, which can help connect motion to function.

How to Use Dyna-1 on Tamarind Bio

1. Log in and open the tool: sign in at tamarind.bio and select Dyna-1.

2. Provide your protein: enter the amino acid sequence, and a structure file if the tool exposes one, since the model uses sequence and structure inputs.

3. Check size: the model was trained on proteins under 400 residues, so shorter single-domain proteins are the best fit.

4. Set options if exposed: use any available parameters on the tool page; the paper does not define user settings beyond the inputs.

5. Run the prediction: submit the job and wait for the per-residue results.

6. Download outputs: retrieve the per-residue probability of exchange, written p(exchange) in the paper.

7. Interpret the results: look for clusters of high-probability residues in loops, binding sites or active sites, and compare with conservation or experimental data.

Things to Keep in Mind

• Training labels assume that missing backbone assignments mainly reflect µs-ms exchange, which the authors call a bold assumption.

• Dyna-1 was trained on proteins under 400 residues, which may limit predictions for larger proteins.

• The model is weakest at predicting exchange in beta-sheets in the authors' analysis.

• It does not predict exchange caused by phosphate binding, and sensitivity to within-family variation and de novo proteins is still under study.

• Predictions are probabilities of exchange, not rates, timescales or structures of the alternative states.

Source: Wayment-Steele, H. K., El Nesr, G., Hettiarachchi, R., Kariyawasam, H., Ovchinnikov, S. and Kern, D., "Learning millisecond protein dynamics from what is missing in NMR spectra," bioRxiv preprint, March 2025, doi:10.1101/2025.03.19.642801. Code: github.com/WaymentSteeleLab/Dyna-1 (noncommercial use).