Cryo-EM structure of an active bacterial TIR–STING filament complex

1Activation of STING signalling results in assembly of oligomeric filament structures that have been observed in both human innate immunity and bacterial antiphage defence–. Key early findings supporting STING oligomerization as a required step of activation include observation of STING puncta formation in cells, electrophoresis analysis of STING multimeric complexes, and artificial activation of STING on fusion to multimerization domains. More recently, insight into the structural basis of mammalian STING oligomerization has been obtained through analysis of STING crystal packing and cryogenic electron microscopy (cryo-EM) structures of tetrameric STING complexes,. Strict conservation of filament formation in prokaryotic STING signalling suggests that prokaryotic and metazoan STING signalling domains share an ancient mechanism of signal induction. To define the molecular basis of STING filament formation, we reconstituted signalling of the S. faecium TIR–STING (Sf STING) antiphage effector in vitro and used single-particle cryo-EM to determine the structure of the activated complex (Fig. 1a and Extended Data Figs. 1, 2, 3 and 8). In response to the nucleotide second messenger c-di-GMP produced during cyclic oligonucleotide-based antiphage signalling system (CBASS) immunity, Sf STING rapidly assembles into oligomers that form single filaments and antiparallel double-filament structures that make supra-molecular contacts between STING and TIR domains of opposing filaments (Extended Data Figs. 1 and 2). The TIR domains are not as well resolved in the main double-filament class, probably owing to conformational heterogeneity, and we therefore focused structural analysis on the single-fibre filaments. A 3.3-Å-resolution cryo-EM reconstruction of the dominant class of single-fibre filaments reveals that Sf STING oligomerizes through formation of a repeating laterally translated array of parallel stacked protein dimers that buries more than 3,000 Å2 of surface area between two pairs of dimers and locks the STING cyclic dinucleotide (CDN)-binding domain and associated TIR effector domains into filamentous assemblies capable of reaching greater than 300 nm in length (Fig. 1a).

2To explain how CDN signal recognition drives filament formation, we determined the cryo-EM structure of Sf STING bound to the weakly activating ligand 3′,3′-cGAMP (ref.; Extended Data Figs. 2 and 3). Compared with the modelled open apo state and partially closed 3′,3′-cGAMP-bound conformations, recognition of the signal from the correct nucleotide, c-di-GMP, induces an inward rotation of the Sf STING β-strand ‘lid’ region of about 25° and about 9° respectively and results in formation of a tightly closed complex (Fig. 1b and Extended Data Fig. 4). Tighter lid closure is driven by repositioning of the highly conserved Sf STING lid domain residue R234 to form base-specific contacts with the guanosine Hoogsteen edge, mirroring interactions required for high-affinity complex formation between human STING and the cGAS product 2′,3′-cGAMP (refs.,; Fig. 1c). The partially closed 3′,3′-cGAMP-bound Sf STING conformation allows incomplete oligomerization and is incompatible with stable filamentous packing beyond about 4–6 units (or about 20 nm). Compared to the fully active Sf STING–c-di-GMP filament structure, 3′,3′-cGAMP recognition induces partial closure of the lid domain and an overall conformation of the STING CDN-binding domain (CBD) that probably weakens contact sites observed for c-di-GMP-induced filaments. Additionally, a lack of well-defined density for the TIR domains in the 3′,3′-cGAMP filaments suggests conformational flexibility that may impact stability of the filaments. Sf STING binds to c-di-GMP with about 300 nM apparent affinity (Kd) and a low nanomolar concentration of c-di-GMP is sufficient to initiate robust NADase activity. Our previous results demonstrate that 3′,3′-cGAMP binds with slightly weaker affinity (about 700 nM Kd) and is unable to induce robust TIR NADase activity, findings that are now explained by our cryo-EM analysis. Thus, complete closure of the lid domain and structural compression around the high-affinity ligand c-di-GMP is essential to translate CDN recognition into a conformation sufficient to seed STING protein-filament formation and downstream signal induction.

3In the Sf STING–c-di-GMP complex, individual dimeric units adopt a two-fold symmetric conformation and form the basic repeating building block of the filament structure (Fig. 1a, right). In each Sf STING dimer unit, the canonical V-shaped STING CBD is positioned above two TIR enzymatic NADase domains that dock against the base of the receptor (Fig. 1a and Extended Data Figs. 4 and 5). The Sf STING CBD contains a unique β-hairpin insertion immediately following the stem dimerization helix, but otherwise adopts the same minimized fold and highly conserved CDN-binding pocket previously observed in crystal structures of Flavobacteriaceae sp. and Capnocytophaga granulosa bacterial STING (ref.). In the active-state Sf STING filament structure, a short linker sequence connects the α-helix stem of each STING domain to the TIR effector domain (Extended Data Fig. 4e). Previously, structures of a TIR–STING homologue from the oyster Crassostrea gigas and the human transmembrane domain-containing STING in inactive states revealed a twisted linker sequence that connects the effector domain to the STING domain located across the dimeric interface,, (Extended Data Fig. 4e). The active-state conformation of the chicken Gallus gallus STING–2′,3′-cGAMP complex exhibits a parallel linker orientation similar to the Sf STING filament, suggesting that parallel linker orientation is a defining feature of both prokaryotic and metazoan STING activation,.

4TIRs are widespread NADase effector domains encoded in CBASS, Pycsar and Thoeris antiphage defence systems,,, but no previous TIR active-state structures exist to explain the mechanism of NAD+ hydrolysis. The SfSTING TIR domain is most closely related to plant immune proteins and secreted bacterial effectors that catalyse glycosidic bond hydrolysis during immune defence and interspecies conflict–. The Sf STING residues F83, F85, L87 and L89 within the highly conserved TIR helix αC interface form extensive hydrophobic packing interactions that bridge the dimeric unit, and the Sf STING TIR domain also contains a unique βD′ and βE′ strand insertion in the TIR ‘CC loop’ that further extends the dimer interface (Fig. 2a–c and Extended Data Fig. 5). Structural comparison with the human SARM1 TIR–ribose structure demonstrates that the SfSTING NAD+-binding pocket is formed by two regions: a set of hydrophobic residues, F6, W33, F37 and L47, positioned to stack the substrate nicotinamide; and a set of hydrophilic residues, S10, R78 and N80, positioned to coordinate the phosphodiester linkage and ribose of the adenosine base (Fig. 2d). In addition to the highly conserved Sf STING catalytic residue E84, the NADase active site is completed by residue D110 from the opposite TIR dimer mate (Fig. 2d and Extended Data Fig. 5).

5The active Sf STING–c-di-GMP structure reveals a series of protein–protein interfaces that explain a shared mechanism of STING filament formation. The primary STING filament interface occurs along two surfaces that pack between adjacent Sf STING dimeric units and drive lateral head-to-head oligomerization (Fig. 3a and Extended Data Fig. 6). These surfaces centred around the hydrophobic residues V280and A309 are exposed in the closed Sf STING domain conformation, explaining a key mechanism that couples c-di-GMP recognition and filament nucleation (Extended Data Fig. 6). Individual STING-domain protomers (STINGa and STINGb) within the Sf STING filament are also bridged by an electrostatic interaction between STINGa R307 and STINGb E290 (Fig. 3b). Notably, the previous cryo-EM structure of a chicken STING–2′,3′-cGAMP tetramer contains residues Q278 and D280 involved in a similar interaction and hydrophobic surfaces packed along the same STING–STING protein interface, revealing remarkable conservation of an ancient mechanism of STING oligomerization (Fig. 3a,b). In the full Sf STING filament structure, the STING domains are more tightly packed compared to those in the minimal chicken/human STING tetramer model. Additionally, a modest approximately 2° shift between packed Sf STING dimeric units is observed in both the single-fibre and wrapped double-fibre cryo-EM reconstructions, resulting in the active Sf STING filament structure adopting a slight curve (Extended Data Fig. 1). Assembly of the complete Sf STING filament allows formation of a second cross-filament interface between the STING domain residues N278 and Q279 and residue E95 in the TIR domain associated with the adjacent protomer (Fig. 3c and Extended Data Fig. 6). Cross-filament TIRa–TIRb interactions are also formed between two flexible Sf STING TIR loops (BB loop: P32a–G43a; and DD loop: A101b–K118b) that stack on top of one another (Fig. 3d). Comparison of the active SfSTING–c-di-GMP complex with the inactive apo C. granulosa bacterial STING structure reveals substantial rearrangements in the STING domain necessary to enable reorganization and TIR-domain packing (Extended Data Fig. 4). Close packing is required to allow the TIR D110 loop to reach across and complete the dimer-mate active site, providing an explanation for how SfSTING filament formation triggers NADase domain activation (Figs. 2d and 3d and Extended Data Fig. 5).

6We next combined the bacterial STING filament structure with biochemical and cellular analysis of SfSTING function to establish a molecular model of STING activation. Measuring degradation of a fluorescent NAD+ analogue, we observed that SfSTING alterations predicted to disrupt STING–STING, STING–TIR and TIR–TIR cross-filament interaction surfaces each strongly inhibit NADase enzymatic activity in vitro (Fig. 4a and Extended Data Fig. 7). The SfSTING substitutions V280D, E290K and R307E within the STING oligomerization interface disrupted all detectable NAD+ hydrolysis. Likewise, SfSTING variants with substitutions in the STING–STING interface (N208D and A309R) and STING–TIR interface (R52E, K142D, N278E, Q279E and D285K) exhibit weak NADase activity only at 10–100× protein concentration, suggesting defects in the ability to oligomerize and catalyse NAD+ cleavage (Fig. 4a and Extended Data Fig. 7). Each SfSTING filament interface mutant retains the ability to form a stable, high-affinity complex with c-di-GMP, demonstrating that inhibition of NADase function is not due to impaired protein stability or ligand interaction (Extended Data Fig. 7 and Supplementary Fig. 1). Negative-stain electron microscopy analysis confirmed that the absence of NADase activity is a direct result of SfSTING interface mutants specifically losing the ability to form an active filament complex (Fig. 4b and Extended Data Fig. 7). Replacement of the TIR BB loop within the principal TIRa–TIRb interaction site with a glycine linker sequence (ΔA36–K41) resulted in complete disruption of NADase function (Fig. 4a and Extended Data Fig. 7). However, this SfSTING TIR mutant retains the ability to oligomerize into a filament in the presence of c-di-GMP, demonstrating that STING–STING interactions are the main driver of filamentation and that secondary TIR–TIR cross-filament interactions are required only for induction of NADase catalysis. We also observe that a D110A mutant retains the ability to recognize c-di-GMP and form long protein filaments but loses all ability to initiate NADase activity, providing further evidence for the essential role of filament formation in the activation mechanism of SfSTING (Extended Data Fig. 7). We expressed SfSTING in Escherichia coli, a bacterium that constitutively produces the activating ligand c-di-GMP, and confirmed that each SfSTING filament interaction interface is essential for STING-induced growth arrest in vivo (Fig. 4c and Extended Data Fig. 7).

7Our results provide a complete structural model of bacterial STING filament formation and effector domain activation in CBASS antiphage defence (Fig. 4d). STING-mediated antiphage defence begins when the associated CBASS protein CdnE, a cGAS/DncV-like nucleotidyltransferase that recognizes a yet unknown phage cue, senses bacteriophage infection and initiates synthesis of the antiviral nucleotide second messenger c-di-GMP (refs.,,). c-di-GMP is a high-affinity ligand that binds STING in a central chamber formed at the receptor homodimeric interface. Cognate CDN signal recognition induces a conformational change in the STING β-strand lid domain that envelopes c-di-GMP in a closed receptor complex. Next, the conformational change induced on complete lid closure exposes surface contact sites that create an interface for nucleating STING filament formation. STING filament extension is driven primarily by STING–STING contacts and cross-filament contacts between STING and the associated TIR effector domain. Finally, filament assembly leads to TIR–TIR interactions that rearrange the NADase active site to stimulate NAD+ degradation and an abortive infection response that prevents phage propagation. In further support of our model of TIR NADase activation in CBASS antiphage defence, another study has determined the high-resolution structure of a distinct CBASS effector named TIR-SAVED that demonstrates that cyclic oligonucleotide binding induces a curved protein filament responsible for TIR NADase activation. In animal cells, protein oligomerization has emerged as a general principle controlling rapid induction of innate immune signalling. Our structural analysis of bacterial STING activation defines the molecular basis of STING filament formation and demonstrates remarkable conservation of oligomerization as a unifying mechanism controlling both prokaryotic and metazoan antiviral immune defence.

8Synthetic CDN ligands were purchased from Biolog Life Science Institute: c-di-GMP (catalogue number C 057) and 3′,3′-cGAMP (catalogue number C 117). Benzamide adenine dinucleotide was a gift from Frank Schwede (Biolog Life Science Institute).

9Recombinant bacterial Sf STING protein was recombinantly expressed and purified as previously described. Briefly, all constructs were cloned using Gibson assembly into a modified pET16 vector for expression of recombinant amino-terminal 6×His-fusion proteins in BL21-CodonPlus(DE3)-RIL E. coli (Agilent). The TIR-to-TIR cross-filament contact mutant ΔA36–K41 was designed as a glycine-serine loop replacement (D35-GSGG-S42). Inoculated 1-l M9ZB cultures (0.5% glycerol, 1% Cas-amino acids, 47.8 mM Na2HPO4, 22 mM KH2PO4, 18.7 mM NH4Cl, 85.6 mM NaCl, 2 mM MgSO4 and trace metals, supplemented with 30 mM nicotinamide to limit TIR toxicity) were grown at 37 °C with 230 r.p.m. shaking. Cultures reaching an optical density at 600 nm (OD600nm) > 2.5 were induced with a final IPTG concentration of 500 μM and incubated at 16 °C overnight at 230 r.p.m. Collected bacterial pellets were sonicated in lysis buffer (20 mM HEPES-KOH pH 7.5, 400 mM NaCl, 30 mM imidazole, 10% glycerol and 1 mM dithiothreitol) and purified by gravity flow over Ni-NTA resin (Qiagen). Resin was washed once with lysis buffer supplemented to 1 M NaCl, and recombinant protein was eluted with 300 mM imidazole. Protein was dialysed overnight at 4 °C (20 mM HEPES-KOH pH 7.5, 250 mM KCl, 10% glycerol and 1 mM dithiothreitol). Dialysed protein was concentrated with 30-kDa-cutoff Amicon centrifuge filters (Millipore) before loading onto a 16/600 Superdex 200 size-exclusion column (Cytiva) equilibrated in gel filtration buffer (20 mM HEPES-KOH pH 7.5, 250 mM KCl, 1 mM TCEP). Protein purity was assessed by denaturing gel before concentrating samples to >10 mg ml−1 and flash freezing in liquid nitrogen for storage at −80 °C.

10On exposure to the activating ligand c-di-GMP, solutions of purified Sf STING immediately begin filament formation and become visibly cloudy. For the first c-di-GMP dataset, SfSTING at 1 mg ml−1 was rapidly mixed with a 3× molar concentration of c-di-GMP (84 μM), immediately applied to glow-discharged 1.2/1.3 Cu 300 mesh grids (Quantifoil), and frozen in liquid ethane within 10 s of mixing using a Vitrobot Mark IV (Thermo Fisher) set at 4 °C and 100% humidity with no wait time, 3 s blot time and +8 blot force. For the second c-di-GMP dataset, SfSTING at 1 mg ml−1 was pre-incubated with 1 mM benzamide adenine dinucleotide before rapid mixing with 84 μM c-di-GMP and frozen as above. Semi-automated data collection was performed with SerialEM v3.8.5 and v3.8.6. Grids were imaged on a Titan Krios (Thermo Fisher) operating at 300 kV equipped with a BioQuantum K3 imaging filter with a 20-eV slit width and a K3 summit direct electron detector (Gatan) in counting mode at a nominal magnification of 105,000× corresponding to a calibrated pixel size of 0.825 Å. For the first dataset, a total exposure time of 1.6 s, corresponding to a total dose of 55.5 electrons Å−2, was fractionated over 49 frames. For the second dataset, a total exposure time of 1.29 s, corresponding to 51.7 electrons Å−2, was fractionated over 51 frames. The defocus targets were −1.2 to −2.1 μm for the first dataset and −1.2 to −2.5 μm for the second dataset.

11For the 3′,3′-cGAMP dataset, SfSTING at 1 mg ml−1 was rapidly mixed with a 3× molar concentration of 3′,3′-cGAMP (84 μM) and frozen as described above. The 3′,3′-cGAMP dataset was collected on a Talos Arctica (Thermo Fisher) operating at 200 kV equipped with a K3 direct electron detector (Gatan) in counting mode at a nominal magnification of 36,000× corresponding to a calibrated pixel size of 1.1 Å. A total exposure time of 4.494 s, corresponding to a total dose of 52.9 electrons Å−2, was fractionated into 50 frames. The defocus targets were −1.4 to −2.6 μm.

12Data processing was performed in cryoSPARC v3.1.018 (ref.) and RELION-3.1 (ref.). For the c-di-GMP datasets, patch-based motion correction and CTF estimation was performed in cryoSPARC. Micrographs with severe contamination or poor contrast transfer function (CTF) fits were removed. Automated particle picking was performed in cryoSPARC with the template picker, using templates generated from either the filament tracer (first dataset) or the blob-based picker (second dataset). The particles were extracted with a box size of 320 and downsampled to a box size of 160 for initial two-dimensional (2D) classification and refinement steps.

13For the c-di-GMP-bound SfSTING single-filament reconstruction, particle coordinates from the filament tracer and template-based picking in the first c-di-GMP dataset were combined, and duplicate coordinates closer than 40 Å were removed. A total of 277,287 coordinates corresponding to single-filament classes after heterogeneous refinement were imported into RELION. The second dataset had more bundled filaments and did not contribute to the single-filament reconstruction. Global and local (12 × 8 patches) motion correction was repeated in RELION using MotionCor2 v1.4.0 (ref.), followed by CTF estimation with GCTF v1.06 (ref.). After 2D classification and 3D refinement, 270,695 particles were subjected to signal subtraction using a mask around the central filament, followed by 3D classification without alignment. A total of 206,965 particles were reverted and subjected to two rounds of CTF refinement and a round of Bayesian polishing. One 3D classification without alignments was performed with the polished particles using a mask around the central filament. A class containing 26,447 particles that best resolved both the TIR and STING domains was selected for a final round of 3D refinement. In our analysis, SfSTING activation is observed as individual filaments that range in size with some filaments reaching >300 nm in length (about 85 dimer copies, about 6.3 MDa). Particles selected for processing and high-resolution structural analysis include density for at least 5 SfSTING dimer copies.

14For the c-di-GMP-bound Sf STING double-filament reconstruction, multiple rounds of heterogeneous refinement were performed independently on each dataset in cryoSPARC to isolate particles contributing to the best reconstructions of a double filament after 3D non-uniform refinement. The final reconstructions contained 176,549 particles in the first dataset and 178,579 particles in the second dataset. As no further density corresponding to the benzamide adenine dinucleotide analogue or other differences were observed in the maps, the double-filament particle coordinates from the two datasets were combined and subjected to local motion correction, CTF refinement and non-uniform refinement. For all datasets, attempts to apply symmetry or helical parameters resulted in inferior reconstructions because the SfSTING dimers are not exactly symmetrically related in the oligomeric complexes.

15For the 3′,3′-cGAMP dataset, patch-based motion correction and CTF estimation was performed in cryoSPARC. Micrographs with severe contamination or poor CTF fits were removed. Automated particle picking was performed in cryoSPARC with the template picker using templates generated from the blob-based picker. The particles were extracted with a box size of 280 and subjected to 2D classification followed by ab initio reconstruction and 3D non-uniform refinement. The resulting map and corresponding 261,685 particle coordinates were exported to RELION. Global and local (12 × 8 patches) motion correction and CTF estimation was repeated in RELION using MotionCor2 and GCTF respectively. After a round of 3D classification, 105,567 particles in classes with clear density for all four strands were subjected to CTF refinement, Bayesian polishing and 3D refinement without and with a mask around the two most defined strands.

16The FsSTING (PDB 6WT5) CBD was used as a starting model docked into the single-fibre c-di-GMP-bound SfSTING density in Coot followed by iterative manual model building. The c-di-GMP-bound SfSTING dimer was used as the starting model in the c-di-GMP-bound double filament and 3′,3′-cGAMP-bound oligomer. In the c-di-GMP double filament, individual secondary structure elements of the TIR domains of the central dimer that interacts with the STING domain of the other filament were placed by rigid fitting and manually adjusted in Coot. N-terminal portions of the TIR domain where side chains were not visible were converted to polyalanine. The TIR domains of all other SfSTING dimers in the c-di-GMP double filament and 3′,3′-cGAMP oligomer were removed. The 3′,3′-cGAMP-bound SfSTING dimers probably contain a combination of the 3′,3′-cGAMP orientation modelled and an approximately 180° rotation. Multiple rounds of Phenix real-space refine was applied with manual correction in Coot in between. Model validation was performed in Phenix using MolProbity (ref.). Figure panels were generated using ChimeraX (ref.) and PYMOL (v2.5.1). Software for data processing and modelling was configured in part by SBGrid (ref.).

17Plate reader reactions to assess NADase function were prepared as described previously. Reactions were built in 50 μl final volume with reaction buffer (20 mM HEPES-KOH pH 7.5, 100 mM KCl), 500 μM nicotinamide 1,N6-ethenoadenine dinucleotide; (ε-NAD, Sigma), 0.1–10 μM enzyme and 20 μM c-di-GMP. Reactions were prepared as master mixes in PCR-tube strips and initiated by adding nicotinamide 1,N6-ethenoadenine dinucleotide immediately before placing into the plate reader. Fluorescence emission at 410 nm was read continuously over 40 min using a Synergy H1 Hybrid Multi-Mode Reader (BioTek) after excitation at 300 nm. Plots were generated with GraphPad Prism 9.3.0.

18SfSTING interactions with radiolabelled c-di-GMP were monitored by electrophoretic mobility shift assay as previously described. In brief, 10-μl reactions contained 1× buffer (5 mM Mg(OAc)2, 50 mM Tris-HCl pH 7.5, 50 mM KCl) with a final protein concentration of 20 μM and about 1 μM α32-P-labelled c-di-GMP generated by overnight reaction of purified Vibrio cholerae DncV with GTP (about 0.1 μCi). Reactions were incubated for 5 min at 25 °C and separated on a 6% nondenaturing polyacrylamide gel held at 100 V for 45 min in 0.5× TBE buffer. Gels were fixed (40% ethanol and 10% glacial acetic acid) before drying at 80 °C for 1 h. Dried gels were then exposed to a phosphor storage screen and imaged on a Typhoon Trio Variable Mode Imager (GE Healthcare).

19Wild-type or mutant Sf STING (1 μM) was incubated with 10 μM c-di-GMP in buffer (20 mM HEPES-KOH pH 7.5, 250 mM KCl, 1 mM TCEP) for 15 min on ice. The mixture was then directly applied to a glow-discharged (30 s, 30 mA) 400-mesh Cu grid (Electron Microscopy Sciences, EMS-400Cu) coated with an approximately 10-nm layer of continuous carbon (Safematic CCU-010) for 30 s. After side blotting, the grid was immediately stained with 1.5% uranyl formate and then blotted again from the side. Staining was repeated twice with a 30-s incubation with uranyl formate in the final staining step. EM images were collected on a FEI Tecnai T12 microscope operating at 120 keV and equipped with a Gatan 4K × 4K CCD camera at a nominal magnification of 52,000× corresponding to a pixel size of 2.13 Å and at a defocus of about 1 μm.

20SfSTING and mutant constructs as well as an sfGFP negative-control construct were cloned into pET vectors for IPTG-inducible expression. E. coli BL21 (DE3) (NEB) were transformed with these plasmids and then plated on LB medium plates supplemented with 100 μg ml−1 ampicillin. After overnight incubation, three colonies from these plates were used to inoculate 5-ml MDG liquid cultures (0.5% glucose, 25 mM Na2HPO4, 25 mM KH2PO4, 50 mM NH4Cl, 5 mM Na2SO4, 2 mM MgSO4, 0.25% aspartic acid and trace metals) supplemented with 100 μg ml−1 ampicillin and grown overnight at 37 °C with 230 r.p.m. shaking. Cultures were diluted 1:50 into fresh M9ZB medium (supplemented with 100 μg ml−1 ampicillin) and grown for 3 h at 37 °C with 230 r.p.m. shaking. Cultures were then diluted to a uniform OD600nm in M9ZB medium and further diluted 1:5 into fresh M9ZB medium supplemented with 5 μM IPTG to induce protein expression. A 200 μl volume of induced culture was added to a 96-well plate in technical triplicate and OD600nm was recorded over 300 min in a Synergy H1 Hybrid Multi-Mode Plate Reader (BioTek) shaking at 37 °C. Plots were generated with GraphPad Prism 9.3.0.