Dynamics linked to biological function—such as enzyme catalysis and ligand binding—are particularly well predicted by the top model, Dyna-1, and residues exhibiting microsecond-to-millisecond exchange tend to be more evolutionarily conserved.

A large-scale observation shows that residues missing in BMRB NMR chemical shift datasets may reflect microsecond-to-millisecond dynamics due to exchange broadening, not merely data gaps.

The work is presented as an unedited Nature manuscript preview, with author affiliations spanning Scripps Research, Stanford, MIT, Harvard, and Carnegie Mellon, and correspondence directed to Dorothee Kern.

A new study curates more than 100 NMR relaxation datasets and trains deep learning models to predict missing NMR assignments, finding these models can also predict exchange measurements from relaxation experiments, signaling they capture microsecond-to-millisecond dynamics.

The work, received in 2025 and accepted on August 3, 2026, was published August 10, 2026, and is accessible via Nature and institutional channels.

The best model, named Dyna-1, incorporates an intermediate layer from the multimodal language model ESM-32 to forecast dynamics-related missing assignments.

The authors propose that the introduced datasets and models could transform understanding by creating a common language that connects protein dynamics with function, enabling broader insights from existing NMR data.