Abstract

Structural dynamics are fundamental to protein functions and mutation effects. Currentprotein deep learning models are predominantly trained on sequence and/or static struc-ture data, which often fail to capture the dynamic nature of proteins. To address this, weintroduce SeqDance and ESMDance, two protein language models trained on dynamicbiophysical properties derived from molecular dynamics simulations and normal modeanalyses of over 64,000 proteins. Both models can be directly applied to predict dynamicproperties of unseen ordered and disordered proteins. SeqDance, trained from scratch,has attentions that capture dynamic interaction and comovement between residues, andits embeddings encode rich representations of protein dynamics that can be further uti-lized to predict conformational properties beyond the training tasks via transfer learning.SeqDance predicted dynamic property changes reflect mutation effect on protein foldingstability. ESMDance, built upon ESM2 (Evolutionary Scale Model II) outputs, substan-tially outperforms ESM2 in zero-shot prediction of mutation effects for designed andviral proteins which lack evolutionary information. Together, SeqDance and ESMDanceoffer a framework for integrating protein dynamics into language models, enabling moregeneralizable predictions of protein behavior and mutation effects.