BPS2026 – Graph neural networks for exploring complex biological mechanisms

Graph neural networks (GNNs) have emerged as a transformative machine-learning approach for molecular dynamics (MD), offering data-driven models that complement traditional physics-based methods. By encoding biomolecules as graphs—where atoms or residues serve as nodes enriched with structural and physicochemical features and edges describe spatial or energetic couplings—GNNs provide a natural representation of chemical space. This representation enables quantum-accurate neural force fields, automated discovery of collective variables to enhance sampling, and efficient prediction of forces to extend simulation reach. Furthermore, GNNs also enable high-level analysis of MD trajectories, using attention mechanisms and transferable graph embeddings to extract mechanistic and interpretable insight. Here, we present a GNN-based strategy to extract mechanistic understanding directly from MD trajectories of large nucleoprotein complexes. We first apply a graph attention network (GAT) to the transposon-associated AAA+ ATPase TnsC, resolving directional filament polymerization along DNA. Edge attention pinpoints residues that bias assembly toward the 3′→5′ direction and rationalizes filament organization. We then extend the same GAT framework to Cas13 in an unsupervised setting, learning to predict dynamics from MD trajectories. Attention maps reveal how distal residues modulate active-site reorganizations required for RNA cleavage. Together, these studies demonstrate that GNN architectures can uncover collective motions and long-range allostery beyond conventional dimensionality reduction or static contact networks, providing an AI-driven, interpretable route to understand complex biological mechanisms.