From Static to Dynamic: Inferring Protein Dynamics from Structure and Language Embeddings

The dynamic behavior of proteins is essential for biological functions such as enzyme catalysis, transmembrane transport, and signal transduction, but its characterization conventionally relies on expensive Molecular Dynamics (MD) simulations. To address this challenge, we develop ResAxial, a deep learning architecture that combines a Residual U-Net with Axial Attention to predict dynamic properties of proteins. We evaluate our model across three input regimes: structure-only (using pairwise distance matrices alone), sequence-only (using ESM-2 embeddings alone), and combined (geometry with sequence embeddings). When combining both modalities with joint supervision on Root Mean Square Fluctuation (RMSF) and correlations, ResAxial achieves strong performance with Correlation PCC=0.959 and RMSF PCC=0.917.