STING's two activation routes diverge in dynamics: ligand binding clusters fluctuations at the lid, SAVI mutations reorganize around a1b1
Synopsis
Molecular dynamics trajectories no longer have to be read only through per-residue fluctuations. Low-population, state-dependent coupling signals were previously hard to separate from high-dimensional trajectories. After compressing each residue pair's side-chain heavy-atom mean-distance time series into a 25-dimensional latent space, then clustering and building networks, the ligand-bound state is dominated by lid residue pairs, while the SAVI-associated multi-mutant shifts to an a1b1-enriched, cross-protomer organization involving residue 281.
Interpretation
One latent-space residue-pair network workflow yields different organizations across three STING conditions: the ligand-bound state centers on the lid, and the SAVI-associated multi-mutant centers on a1b1. Residues were previously described by per-residue summary metrics such as root-mean-square fluctuation, which struggle to isolate low-population, state-dependent fluctuation patterns. In the d5.5–6.0 Mahalanobis-distance band, the ligand-bound Holo-C1 major connected component has 49 nodes and 192 edges, with 93.9% of nodes containing lid residues; the corresponding Mut-C2 component has 31 nodes and 101 edges, with a1b1, lid, and peripheral-LBD nodes at 54.8%, 41.9%, and 38.7%.
Each residue pair is represented by an 1800-point time series, which the encoder compresses into a low-dimensional vector describing its fluctuation pattern, and nodes are then linked by similarity. Edges encode similarity between learned latent descriptors, not physical inter-residue distances or successive time intervals. The pipeline trains a one-dimensional convolutional autoencoder on apo trajectories, applies temporal mean pooling, and removes seven all-zero dimensions to yield 25-dimensional descriptors; each trajectory exceeds 1 microsecond, and the 100–1000 ns window at 0.5 ns intervals gives 1800 structures and 50,403 residue pairs per state.
In the mutant network, residue 281 sits inside a cross-region subnetwork rather than at a lid-centered position as in the ligand-bound state. In the ligand-bound networks, nodes containing wild-type 281/284 are present, but the highest degree-centrality nodes are mainly lid and neighboring LBD pairs. In Mut-C2 at d5.5–6.0, node 191B_281B and its six directly connected neighbors define the SAVI-associated subnetwork; the top hub in that band is 303A_305B with 13 connections and degree centrality 0.325, three of which fall inside that subnetwork.
Perspective
The analysis covers the membrane-embedded STING dimer; explicit CTT engagement and higher-order polymerization lie outside the modeled systems. The multi-mutant combines eight SAVI-associated substitutions to enrich mutation-associated dynamic signatures and is a state-level stress-test model rather than a specific patient genotype. Each condition uses one trajectory exceeding 1 microsecond, so the comparison characterizes the three analyzed datasets.
Clustering was performed separately per state, so the numerically dominant background cluster does not imply one-to-one identity of the same residue pairs across conditions; state-dependent differences concentrate in small, low-population sectors of latent-feature space. The authors note that assessing whether these patterns are consistent across runs would require application to independent trajectories.