Latent-Space Residue-Pair Networks from Molecular Dynamics Distinguish Ligand-Activated and SAVI-Associated STING

Abstract

Stimulator of interferon genes (STING) provides an informative model system for developing methods to extract and compare residue-pair fluctuation patterns from molecular dynamics (MD) trajectories. We combined a one-dimensional convolutional autoencoder trained on apo STING with temporal mean pooling, Mahalanobis-distance clustering, and residue-pair similarity networks to compare apo, ligand-bound, and SAVI-associated multi-mutant datasets using 25-dimensional latent descriptors. Networks constructed from clusters capturing state-dependent fluctuation patterns showed lid-dominated organization in the ligand-bound dataset and a1b1-enriched, structurally distributed relationships involving residue 281 in the multi-mutant dataset. These findings identify contrasting residue-pair fluctuation signatures in the analyzed trajectories and provide an interpretable framework for generating hypotheses about ligand-associated and mutation-associated dynamic allostery.