Deep learning model predicts protein millisecond dynamics from missing NMR data

Protein functions such as enzyme catalysis and ligand binding often depend on conformational changes occurring at micro- to millisecond timescales. Experimental data on these motions has been limited, making it difficult to develop a predictive understanding of protein dynamics.

Researchers curated more than 100 nuclear magnetic resonance (NMR) relaxation datasets and observed that millisecond dynamics can broaden NMR signals beyond detection. This causes some residues to go unassigned in chemical shift datasets for approximately 10,000 proteins deposited in the Biological Magnetic Resonance Data Bank.

The team hypothesized that missing residue assignments indicate exchange broadening from micro- to millisecond motions. They trained several deep learning models to predict which residues would be missing from assignments, using the multimodal language model ESM-3 as a foundation.

The best performing model, named Dyna-1, successfully predicted exchange rates measured independently through NMR relaxation experiments. This agreement suggests the model captures genuine dynamic behavior rather than merely predicting data gaps.

Dyna-1 showed particular accuracy for residues where dynamics are directly linked to biological function, including enzyme active sites and ligand binding regions. The researchers also found that residues experiencing micro- to millisecond exchange tend to be more evolutionarily conserved.

The study provides standardized, large-scale predictions of protein dynamics across thousands of structures. The authors anticipate the datasets and models will help reveal common principles connecting protein motion to function.