ESMDynamic: Fast and accurate prediction of protein dynamic contact maps from single sequences
TL;DR
ESMDynamic is a deep learning model that predicts protein residue-residue contact dynamics directly from single sequences, matching or outperforming state-of-the-art ensemble methods like AlphaFlow while requiring orders-of-magnitude less computation.
Problem / question
Most deep learning models in structural biology predict only static protein structures, leaving a gap in understanding the conformational dynamics, equilibrium populations, and kinetics that are essential for elucidating protein function.
Methods
The researchers developed ESMDynamic, a deep learning model built on the ESMFold architecture, trained on experimental structure ensembles and molecular dynamics (MD) simulations. It predicts dynamic contact probabilities, equilibrium contact frequencies, and coarse-grained kinetics across multiple temperatures, and was benchmarked on large-scale MD datasets (mdCATH and ATLAS) and applied to over 18,000 proteins in the human proteome.
Key findings
ESMDynamic matched or outperformed state-of-the-art ensemble prediction methods (AlphaFlow, ESMFlow, BioEmu) on the mdCATH and ATLAS benchmarks while using orders-of-magnitude less computational power. The model successfully generalized to diverse systems, including membrane transporters, a de novo designed protein, and a homodimer complex. Additionally, the predicted dynamic contacts enabled the automated selection of collective variables for constructing Markov state models, and the tool successfully generated predictions for over 18,000 proteins in the human proteome.
Why it matters
By providing a fast, scalable, sequence-based method to predict protein dynamics, ESMDynamic enables large-scale analysis of conformational variability across entire proteomes and accelerates simulation and design workflows.
Limitations
The provided text does not explicitly state specific limitations, caveats, or biases of the ESMDynamic model.
Takeaway
ESMDynamic accurately and rapidly predicts protein dynamic contact maps and kinetics directly from single sequences, offering a highly scalable alternative to computationally expensive molecular dynamics simulations.