Learning Enhanced Protein Representations from Molecular Dynamics via Temporal GNNs

Proteins are inherently dynamic entities, yet traditional deep learning approaches predominantly rely on static structural representations. In this work, we employ Temporal Graph Neural Networks (Temporal GNNs) to encode the temporal evolution of Molecular Dynamics (MD) trajectories. Using the MISATO dataset, we show that predicting directly from dynamic data outperforms static baselines on atomic adaptability prediction and binding site detection. To address the computational cost of dynamic models, we propose a feature-based transfer learning strategy that distills dynamic knowledge learned through self-supervised pre-training into dense node features. Augmenting static models with these representations improves performance across tasks while avoiding the need to train costly temporal models for each downstream application. Our results suggest that dynamic knowledge captured by MD simulations can be effectively compressed and transferred to static models, bridging the gap between the richness of physics-based simulations and the efficiency required for practical applications.