Skip to content
DraconiticResearch
Search Sources API
197,768 studies31 sourcesUpdated 24 min ago
Peer-reviewedCC BY-NC-NDNature Communications

ESMDynamic: Fast and accurate prediction of protein dynamic contact maps from single sequences

  • Diego E. Kleiman1,2
  • Jiangyan Feng3,2
  • Zhengyuan Xue1,2
  • Diwakar Shukla4,5,6,7,8,2
8 affiliations
  1. 1Center for Biophysics and Quantitative Biology, University of Illinois at Urbana-Champaign, Urbana, IL, USA
  2. 2University of Illinois at Urbana-Champaign
  3. 3Department of Chemical and Biomolecular Engineering, University of Illinois at Urbana-Champaign, Urbana, IL, USA
  4. 4Center for Biophysics and Quantitative Biology, University of Illinois at Urbana-Champaign, Urbana, IL, USA. [email protected]
  5. 5Department of Bioengineering, University of Illinois at Urbana-Champaign, Urbana, IL, USA. [email protected]
  6. 6Department of Chemical and Biomolecular Engineering, University of Illinois at Urbana-Champaign, Urbana, IL, USA. [email protected]
  7. 7Department of Chemistry, University of Illinois at Urbana-Champaign, Urbana, IL, USA. [email protected]
  8. 8Department of Plant Biology, University of Illinois at Urbana-Champaign, Urbana, IL, USA. [email protected]
Open at sourceDOICite

Abstract

Understanding conformational dynamics is essential for elucidating protein function, yet most deep learning models in structural biology predict only static structures. Here, we present ESMDynamic, a deep learning model that predicts residue-residue contact dynamics directly from protein sequence. Built on the ESMFold architecture and trained on conformational variability from experimental structure ensembles and molecular dynamics (MD) simulations, ESMDynamic predicts dynamic contact probabilities, contact occupancy fraction, and coarse-grained kinetics of contact formation and dissociation across multiple temperature conditions. On large-scale MD benchmarks (mdCATH and ATLAS), ESMDynamic matches or outperforms state-of-the-art ensemble prediction methods (AlphaFlow, ESMFlow, BioEmu) while requiring orders-of-magnitude less computation. We demonstrate generalization to diverse systems, including membrane transporters, a de novo designed protein, and a homodimer complex. We show that predicted dynamic contacts enable automated selection of collective variables for Markov state model construction. Applied to the human proteome, ESMDynamic generates predictions for over 18,000 proteins, enabling large-scale analysis of conformational variability. Overall, ESMDynamic provides a scalable, sequence-based representation of protein dynamics to inform simulation, analysis, and design workflows.

Related

  1. Accurate and efficient prediction of protein conformations with ProtMonomer
    Y. Si, S. Zhang, L. Chen
    bioRxivPreprintCC BY-NDFull textMolecular biology & gene regulationbioRxiv
  2. Towards Accurate Prediction of Mutation-Induced Changes in Protein Structure
    Zhuoyi Liu, Alex Calabrese, Corey S. O'Hern
    arXivPreprintarXiv licenceFull textBiophysicsBiological & medical physicsarXiv
  3. Accurate quantification of protein concentrations and state populations by 2D 1H-N Q-BEST NMR
    Yacine Ouali, Beate Bersch, Bernhard Brutscher
    Journal of Biomolecular NMRPeer-reviewedCC BY 4.0Full textBiophysicsStructural biologyHALarXiv
  4. Accurate prediction of CRISPR editing outcomes in somatic cell lines and zygotes with few-shot learning with inDecay.
    Weizhong Zheng, Lu Yu, Chao Wu +14
    Genome biologyPeer-reviewedCC BY-NC-NDFull textMolecular biology & gene regulationEurope PMCDOAJ
  5. Accurate structural modelling of protein-glycan recognition with GlycoWeave
    Huifeng Zhao, Jintu Zhang, Yifei Xiao +18
    arXivPreprintCC BY-NC-SA 4.0Full textBiophysicsarXiv

Cite

BibTeXRIS
BibTeX
@article{kleiman2026esmdynamic,
  title = {{ESMDynamic: Fast and accurate prediction of protein dynamic contact maps from single sequences}},
  author = {Diego E. Kleiman and Jiangyan Feng and Zhengyuan Xue and Diwakar Shukla},
  journal = {Nature Communications},
  year = {2026},
  month = aug,
  doi = {10.1038/s41467-026-76361-2},
  url = {https://doi.org/10.1038/s41467-026-76361-2},
  note = {CC BY-NC-ND},
}
RIS
TY  - JOUR
TI  - ESMDynamic: Fast and accurate prediction of protein dynamic contact maps from single sequences
AU  - Diego E. Kleiman
AU  - Jiangyan Feng
AU  - Zhengyuan Xue
AU  - Diwakar Shukla
PY  - 2026
DA  - 2026/08/11
JO  - Nature Communications
DO  - 10.1038/s41467-026-76361-2
UR  - https://doi.org/10.1038/s41467-026-76361-2
AB  - Understanding conformational dynamics is essential for elucidating protein function, yet most deep learning models in structural biology predict only static structures. Here, we present ESMDynamic, a deep learning model that predicts residue-residue contact dynamics directly from protein sequence. Built on the ESMFold architecture and trained on conformational variability from experimental structure ensembles and molecular dynamics (MD) simulations, ESMDynamic predicts dynamic contact probabilities, contact occupancy fraction, and coarse-grained kinetics of contact formation and dissociation across multiple temperature conditions. On large-scale MD benchmarks (mdCATH and ATLAS), ESMDynamic matches or outperforms state-of-the-art ensemble prediction methods (AlphaFlow, ESMFlow, BioEmu) while requiring orders-of-magnitude less computation. We demonstrate generalization to diverse systems, including membrane transporters, a de novo designed protein, and a homodimer complex. We show that predicted dynamic contacts enable automated selection of collective variables for Markov state model construction. Applied to the human proteome, ESMDynamic generates predictions for over 18,000 proteins, enabling large-scale analysis of conformational variability. Overall, ESMDynamic provides a scalable, sequence-based representation of protein dynamics to inform simulation, analysis, and design workflows.
SN  - 2041-1723
ER  -