ESMDance and SeqDance are two protein language models training on dynamic properties derived from molecular dynamics trajectories of over 35,800 proteins and normal mode analyses of over 28,500 proteins. SeqDance, which operates solely on sequence input, captures local dynamic interactions and global conformation properties for both ordered and disordered proteins, even for proteins without homologs in the pre-training dataset. SeqDance predicted dynamic property changes are predictive of mutation effects on protein folding stability. ESMDance, which utilizes ESM2 outputs, significantly outperforms ESM2 in zero-shot prediction of mutation effects for designed and viral proteins. SeqDance and ESMDance offer novel insights into protein behaviors and mutation effects through the perspective of protein dynamics
