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

Understanding conformational dynamics is essential for elucidating protein function, yet most deep learning models in structural biology predict only static structures. Here, we introduce ESMDynamic, a deep learning model that predicts dynamic residue-residue contact probability maps directly from protein sequence. Built on the ESMFold architecture, ESMDynamic is trained on contact fluctuations from experimental structure ensembles and molecular dynamics (MD) simulations, enabling it to capture diverse modes of structural variability without requiring multiple sequence alignments. We benchmark ESMDynamic on two large-scale MD datasets (mdCATH and ATLAS), showing that it matches or outperforms state-of-the-art ensemble prediction models (AlphaFlow, ESMFlow, BioEmu) for transient contact prediction while offering orders-of-magnitude faster inference. We demonstrate the model on the ASCT2 and SWEET2b transporters, a de novo troponin C design, and the HIV-1 protease homodimer, illustrating generalization to unseen systems and recovery of experimentally validated dynamic contacts. Furthermore, we present an automated pipeline using ESMDynamic predictions to select collective variables for Markov State Model construction, producing high-quality kinetic models from unbiased MD simulations of SWEET2b. Overall, ESMDynamic provides a compact and interpretable sequence-based description of conformational dynamics, with broad applications in protein engineering, functional analysis, and simulation-guided discovery.