Researchers unveil Dyna-1 model predicting protein dynamics from missing NMR data
This digest was compiled by AI from multiple sources — links to the originals are below.
An international research team reports in Nature the development of Dyna-1, a deep learning model that infers protein motions on microsecond-to-millisecond timescales from missing NMR assignments. The model leverages the ESM-3 protein language model and was trained on over 100 curated NMR relaxation datasets, treating unobservable signals as a predictive feature. Dyna-1 accurately forecasts exchange broadening linked to enzyme catalysis and ligand binding, aligning with evolutionary conservation patterns.
The Missing Signals
Chemical shift assignments in the Biological Magnetic Resonance Data Bank (BMRB) are often incomplete, with some residues unobserved due to signal broadening. The team hypothesized that these missing assignments result from exchange broadening caused by microsecond-to-millisecond conformational dynamics. Curating more than 100 NMR relaxation datasets, they confirmed that unassigned residues are a robust proxy for motion, turning a common experimental nuisance into a predictive signal. The BMRB contains approximately 10,000 protein datasets, offering a vast resource for mining dynamic information.
Dyna-1 Model
Dyna-1 is a deep learning model built on an intermediate representation from ESM-3, a multimodal protein language model. Trained to predict missing assignments, the model unexpectedly also forecasts quantitative exchange measurements from relaxation experiments. Dyna-1 excels at identifying dynamics at functional sites: residues involved in enzyme catalysis and ligand binding are predicted with high accuracy. The study further revealed that amino acids experiencing µs-ms exchange are under stronger evolutionary selection, linking dynamics directly to biological importance.
Biological Implications
The researchers anticipate that Dyna-1 and its associated datasets will unlock a new era of understanding how protein dynamics encode function. The model allows scientists to infer motion from existing static chemical shift assignments without additional experiments, potentially revealing general principles of how evolution tunes protein flexibility. By shifting the paradigm from ignoring missing data to actively exploiting it, the work demonstrates a transformative approach to studying the common language of dynamics and function.
What's Next
The team plans to release open-source models and datasets for the community. It remains unclear whether Dyna-1’s predictions will generalize to membrane proteins and intrinsically disordered regions, which also pose challenges for NMR.
1 source
Researchers unveil Dyna-1 model predicting protein dynamics from missing NMR data
