STING's two activation routes diverge in dynamics: ligand binding concentrates fluctuations in the lid, SAVI mutations reorganize them around a1b1

The analysis treats each residue pair as a node and connects nodes whose fluctuation patterns are similar, not residues that touch in space. Earlier descriptions of molecular dynamics mostly summarized each residue with a single metric such as root-mean-square fluctuation, which makes low-population, state-dependent coupling hard to isolate. In the ligand-bound dataset, the networks built from the two selected minor clusters contained lid residues in 93.9% (46 of 49 nodes) and 95.0% (19 of 20 nodes) of their nodes, and the highest-centrality nodes were mainly lid and neighboring ligand-binding-domain residue pairs.

The SAVI-associated multi-mutant does not reproduce the lid-dominated ligand-bound pattern; it redistributes coordinated fluctuations toward the a1b1 region and cross-region connections. In the apo state the two largest minor clusters drew roughly 75% of their pairs from the peripheral ligand-binding domain, whereas more than half of the connected nodes in the mutant Mut-C2 contained a1b1 residues in all three similarity bands. In the d5.5–6.0 band, the major connected component of Mut-C2 contained 31 nodes and 101 edges, 17 of them (54.8%) containing a1b1 residues, and the 281-containing node 191B_281B sat inside that component with six direct neighbors forming the SAVI-associated subnetwork.

The method compresses each residue-pair distance trajectory into a low-dimensional descriptor, clusters those descriptors to isolate a few non-background behaviors, and then builds networks on descriptor similarity. A one-dimensional convolutional autoencoder was trained only on apo trajectories, and the same trained model was then applied to the apo, holo, and mutant states, so all three states share one descriptor coordinate system. Each trajectory extended beyond 1 microsecond, and the 100–1000 ns window yielded 1800 structures per state at 0.5 ns intervals; 50,403 residue pairs per state were clustered into one background cluster of more than 50,000 pairs and nine minor clusters of fewer than 100 pairs each.

The workflow is demonstrated on the membrane-embedded STING dimer across three states, apo, ligand-bound, and a SAVI-associated multi-mutant, using one trajectory beyond 1 microsecond per state and a 100–1000 ns analysis window. The multi-mutant was built by introducing eight SAVI-associated substitutions into the apo coordinates, and the authors describe it as a state-level stress-test model rather than a representation of a specific patient genotype. Explicit C-terminal-tail engagement and higher-order polymerization lie outside the modeled systems. The authors propose that applying the same workflow to independent trajectories would let the consistency of these patterns across runs be assessed.

The authors note that clustering was performed separately for each state, so the background clusters do not imply one-to-one identity at the residue-pair level, and that state-dependent differences are concentrated in small, low-population sectors of latent-feature space. The selected minor clusters are small: 66 and 38 pairs in apo, 96 and 46 in holo, and 82 and 58 in the mutant. Network analysis was restricted to Mahalanobis distances of at most 6.0 because larger distances steeply increased the number of connected node pairs and made the networks hard to interpret. The authors also state that the held-out reconstruction loss contributed to model selection, so it served as a reconstruction criterion rather than an independent estimate of predictive performance.