Answers are grounded in this paper and cite their sources.
Molecular dynamics (MD) simulations provide insights about the molecular world, but only if analyzed properly. Determining the relevant dynamics in a protein–ligand system often becomes the main bottleneck of the analysis. Here, we analyze the performance of graph neural networks (GNNs) trained on the MD simulations of several complexes of the innate immunity receptor TLR4/MD2 with different glycolipids to disentangle the atomic/molecular characteristics that determine the agonist/antagonist activity. By systematically evaluating distinct atomic subselections, we found that the classification performance of our model is higher for backbone‐based subselections compared to all‐atom representations. These findings indicate that dynamical differences in the agonist–antagonist behavior are primarily modulated by collective dynamics of TLR4/MD2 rather than specific ligand–residue interactions located at MD2. After analyzing the connectivity of the nodes along the simulations, MD2‐polar rim and surrounding regions play a different behavior upon the binding of agonist or antagonist glycolipids, indicating a key role in the modulation of TLR4 activity. This work shows how MD simulation analysis through GNNs provides a powerful tool to extract meaningful atomic information to identify functionally relevant binding features in protein‐ligand complexes.
No takes yet. Share an insight, caveat, or question.
Martín-Cámara et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: