Graph Neural Network - State Predictive Information Bottleneck with Enhanced Sampling for Protein Ligand Systems

The observable timescale of a protein-ligand systems' residence time is a difficult feat for molecular dynamics (MD) simulations even with the most capable supercomputers. Recent efforts in rare event enhanced sampling methods have been able to lower this gap though it still remains a challenge. One of those aspects is the need for differentiable collective variables (CVs) that can be known to amplify the rare event outcome. Here, we propose using the Graph Neural Network - State Predictive Information Bottleneck (GNN-SPIB) framework to minimize the need for handcrafted CVs altogether. Using coordinate information for input features, we explore different graph model configurations to observe faster dissociation times.