AI predicts hidden millisecond protein motions from absent NMR data

Scientists inferred micro- to millisecond protein dynamics by training deep-learning models on residues missing from NMR datasets.

A group curated more than a hundred NMR relaxation datasets and observed that numerous proteins deposited in the Biological Magnetic Resonance Data Bank have unassigned residues. They hypothesized that these gaps arise from exchange-broadening caused by conformational exchange on micro- to millisecond timescales. Leveraging this idea, they trained several deep-learning architectures to infer the missing chemical-shift assignments, discovering that the predictions align with exchange rates obtained from conventional NMR relaxation experiments.

Their top model, Dyna-1, incorporates an intermediate layer from the multimodal language model ESM-32 and excels at detecting dynamics linked to functional processes such as catalysis and ligand binding. The study further reports that residues undergoing µs-ms exchange tend to be evolutionarily conserved. The authors anticipate that the curated datasets and predictive models will help bridge protein dynamics with biological function.