DPLM: Dynamics-aware Protein Language Model via contrastive learning between sequence and molecular dynamics simulation trajectory

Protein dynamics captures the interactions between residues and the conformational changes they undergo. Encoding such information could enhance a deep learning model's ability to predict protein properties. Here we introduce a dynamics-aware protein language model (DPLM) that encodes protein dynamics into a sequence-derived representation. We represent each molecular dynamics (MD) trajectory as a sequence of backbone C contact maps and encode it with a Video Vision Transformer (ViViT). We then align the ViViT-derived dynamics embedding with the ESM2-derived sequence embedding using dual-view contrastive learning. The root-mean-square fluctuation correlation and dynamic cross-correlation matrix reconstruction confirm that contrastive learning incorporates dynamics information into the sequence embedding. We applied DPLM as a protein representation backbone for predicting diverse dynamics-related downstream tasks, including liquid-liquid phase separation, cryptic-site detection, and mutation effect on protein thermal stability. DPLM shows strong performance, especially where the evolutionary or structural signal is weak. These results show that learning the geometry of molecular motion yields transferable dynamic information that complements the evolutionary and structural features of existing PLMs.