ESMDynamic: Fast and Accurate Prediction of Protein Dynamic Contact Maps from Single Sequences.

Area of Science:

• Structural Biology
• Computational Biology
• Biophysics

Background:

• Understanding protein conformational dynamics is crucial for function but challenging to predict.
• Current deep learning models often predict static protein structures, neglecting dynamic behavior.

Purpose of the Study:

• To introduce ESMDynamic, a novel deep learning model for predicting dynamic residue-residue contact probability maps directly from protein sequences.
• To enable sequence-based prediction of protein structural variability without multiple sequence alignments.

Main Methods:

• ESMDynamic builds upon the ESMFold architecture.
• The model is trained on contact fluctuations from experimental structure ensembles and molecular dynamics (MD) simulations.
• Benchmarked on mdCATH and ATLAS datasets against state-of-the-art ensemble prediction models.

Main Results:

• ESMDynamic matches or surpasses existing models in predicting transient contacts.
• Achieves orders-of-magnitude faster inference speeds compared to other methods.
• Demonstrates successful application to transporters, designed proteins, and viral proteindimers, recovering validated dynamic contacts.

Conclusions:

• ESMDynamic offers a fast, interpretable, sequence-based method for characterizing protein conformational dynamics.
• The model facilitates the construction of kinetic models from MD simulations.
• Broad applications include protein engineering, functional analysis, and simulation-guided discovery.