Main

NINJ1 is a 16-kDa plasma membrane protein that is evolutionarily conserved and found in all higher eukaryotes. It has been predicted to feature two transmembrane helices, with the termini located in the extracellular space8 (Extended Data Fig. 1a). During inflammasome-driven pyroptosis, NINJ1 induces plasma membrane rupture (PMR) downstream of the cell death executor gasdermin D (GSDMD), which in turn is activated by caspase-dependent cleavage8,9,10 (Extended Data Fig. 1b–f and Supplementary Videos 1 and 2). PMR coincides with the formation of higher-order NINJ1 polymers and membrane blebs8. To better understand the correlation between these events and to study the assembly kinetics of NINJ1 polymers in response to inflammasome activation, we performed crosslinking experiments in mouse bone-marrow-derived macrophages (BMDMs). In line with progressive polymerization of NINJ1, we detected the formation of NINJ1 dimers and trimers 10 min after inflammasome activation; this was followed by extensive polymerization of NINJ1 and the formation of larger polymers at later time points (Fig. 1a,b). The bulk of NINJ1 polymerization coincided with complete oligomerization of cleaved GSDMD p30 (Fig. 1a,b)—that is, formation of GSDMD pores—consistent with NINJ1 activation occurring downstream of GSDMD activation. PMR was quantified by the release of lactate dehydrogenase (LDH), which is too large to be released directly through GSDMD pores11,12,13. Of note, following inflammasome activation, the amount of released LDH increased slowly with time, whereas NINJ1 polymerization was already detectable at the onset of LDH release and increased only marginally at later time points (Fig. 1c). Small amounts of NINJ1 dimers were also detectable in live cells (Fig. 1a,b), suggesting that higher-order NINJ1 polymers are the active species. These time-resolved data are thus fully consistent with a necessity of NINJ1 polymerization for PMR.

Fig. 1: Polymerization kinetics of plasma membrane NINJ1.
figure 1

a,b, Western blot analysis of endogenous GSDMD and NINJ1 in primed BMDMs after nigericin stimulation (Nig) for 1.5 h (a) or for 2, 5, 15, 25, 55 or 85 min (b) followed by treatment with the membrane-impermeable BS3 crosslinker for 5 min. FL, full length. Of note, tubulin, used here as a loading control, is crosslinked owing to BS3 entry through GSDMD pores under nigericin-treated conditions. Time in b is the total incubation time for nigericin and BS3 treatment. Short exp., short exposure. For gel source data, see Supplementary Fig. 1. c, LDH release from primed BMDMs after nigericin stimulation. d, Time-lapse fluorescence confocal microscopy of HeLa cells co-expressing hNINJ1–GFP and opto-casp1 following photo-activation. Images show the NINJ1–GFP fluorescence at the basal plane of the cell and the influx of DRAQ7 (maximum (max.) projection from a z-stack) to track plasma membrane permeabilization. Time was normalized to the onset of increase in DRAQ7 nuclear fluorescence. White arrows indicate regions that are enlarged in the insets. Scale bar, 10 µm. eg, Normalized quantification of the distribution inhomogeneity of NINJ1–GFP (e), HATMD–GFP (f) or E-cadherin–GFP (g) at the basal plane of cells, and DRAQ7 nuclear fluorescence intensity after photoactivation of opto-casp1 (F). The distribution inhomogeneity at each time point (Dt) was normalized to the distribution inhomogeneity at the initial time point of the experiment (Di). Data are mean ± s.d. Data are representative of 2 (b), 3 (a) or 14 (d) independent experiments, or pooled from 2 independent experiments performed in triplicate (c) or at least 10 (eg) independent experiments.

Source data

We also monitored NINJ1 polymerization using time-lapse fluorescence microscopy in HeLa cells co-expressing GFP-tagged human NINJ1 (hNINJ1) and a CRY2–caspase-1 fusion protein (opto-casp1), which enables rapid caspase activation and induction of GSDMD-driven pyroptosis in single cells using optogenetics14 (Fig. 1d and Extended Data Fig. 2a–c). Quantitative inhomogeneity analysis showed that concurring with the influx of the dye DRAQ7, which indicates a loss of plasma membrane integrity and cell death, the diffusely localized NINJ1 started to cluster at the plasma membrane (Fig. 1d,e, Extended Data Fig. 2c and Supplementary Video 3), and that these clusters persisted beyond cell lysis. Notably, the formation of clusters during pyroptosis was specific for NINJ1, as other plasma membrane proteins such as the haemagglutinin transmembrane domain15 (HATMD) or E-cadherin did not cluster (Fig. 1f,g, Extended Data Fig. 2d,e and Supplementary Videos 4 and 5). Similar assemblies were also formed by endogenous NINJ1 in wild-type but not in Gsdmd—/— BMDMs upon activation of the NLRP3 or AIM2 inflammasomes (Extended Data Fig. 3a). Clustering of NINJ1 and NINJ1-dependent LDH release was also detected upon induction of apoptotic cell death8 (Extended Data Fig. 3a,b), but in this case independently of GSDMD. Next, we investigated whether NINJ1-driven PMR is a cell-intrinsic process or whether this activity is dependent on neighbouring cells via released NINJ1 or direct contact. Co-culture experiments of wild-type BMDMs with Casp11-deficient BMDMs transfected with lipopolysaccharide (LPS) to activate the non-canonical inflammasome unambiguously demonstrated that NINJ1 lyses cells in a cell-intrinsic manner without affecting immediate neighbours (Extended Data Fig. 3c). NINJ1-driven PMR in inflammasome-activated cells is thus a cell-intrinsic process, which involves the formation of GSDMD pores and rapid NINJ1 polymerization, followed by membrane rupture with slower kinetics.

NINJ1 polymerizes into filaments

We next performed stochastic optical reconstruction microscopy (STORM) to investigate the nanoscale organization of NINJ1 assemblies in pyroptotic cells. To this end, we used HeLa cells co-expressing hNINJ1–GFP and DmrB–caspase-4. DmrB is a dimerization module that enables activation of caspase-4, the human caspase-11 orthologue, and subsequent activation of GSDMD16,17 by an exogenous, cell-permeable ligand18 (Extended Data Fig. 4a). We used an anti-GFP nanobody and wide-field epi-fluorescence imaging to specifically label and image hNINJ1–GFP structures (Fig. 2a). For STORM, total internal reflection fluorescence (TIRF) illumination with single-molecule localization analysis was used to super-resolve NINJ1 assemblies in the plasma membrane (Fig. 2b). Live cells showed a homogeneous distribution of small NINJ1 dots, which could be individual molecules or small polymers, and a few small and mostly round clusters (Fig. 2b). By contrast, cells undergoing pyroptosis displayed larger NINJ1 clusters, with dimensions ranging from about 500 nm to several micrometres (Fig. 2b,c). The morphology of these clusters was highly branched, with individual branches protruding in different directions, somewhat resembling the organization of the filamentous ASC speck. These clusters correspond to the large dots observed by wide-field epi-fluorescence and confocal microscopy (Fig. 1d and Extended Data Fig. 2c). Again, cluster formation during pyroptosis was not observed for the control protein HATMD (Extended Data Fig. 4b). To quantify the change in NINJ1 nanoscale organization upon activation of cell death, we performed density-based clustering to determine the number of clusters per area, the distributions of cluster sizes (radius of gyration (Rg)) and shapes (eccentricity (Ecc)), and the number of single-molecule localizations per NINJ1 cluster. This analysis showed that after caspase activation, more clusters were identified per cell surface area and the clusters were significantly larger and less spherical than in non-activated cells (Fig. 2d and Extended Data Fig. 4c–f). In around 10% of cells, we also observed long NINJ1 filaments up to several micrometres in length that connected the larger assemblies (Fig. 2c,e). In summary, super-resolution microscopy analysis showed that in pyroptotic cells, NINJ1 polymerizes into large clusters with various shapes of branched, filamentous morphology, as well as long filaments in the micrometre range.

Fig. 2: Super-resolution imaging of NINJ1 assemblies.
figure 2

a, Wide-field imaging of DmrB–Casp4tg HeLa cells expressing hNINJ1–GFP and GSDMD used for STORM microscopy in be. Cells were left untreated or stimulated with B/B homodimerizer 3 h before fixation and labelling with Alexa Fluor 647-conjugated anti-GFP nanobodies. Scale bar, 50 µm. be, STORM super-resolution imaging of hNINJ1–GFP in cells from a using TIRF illumination of the basal plane. b, Left, untreated or B/B-stimulated cells expressing hNINJ1–GFP. Scale bar, 10 µm. b, Right, STORM super-resolution reconstruction of hNINJ1–GFP labelled with Alexa Fluor 647-conjugated anti-GFP nanobodies. The indicated outlined regions are magnified on the right. Scale bar, 500 nm. c, Gallery of hNINJ1–GFP clusters found in pyroptotic cells. The small clusters are also observed in non-activated cells. Scale bars, 500 nm. d, Radius of gyration (Rg) and eccentricity (Ecc) for each identified hNINJ1 cluster. Plots show the distribution of all identified clusters from three independent experiments. The lines indicate median values. Statistical analysis based on the median Rg and Ecc of each experiment using Student’s unpaired two-sided t-test. **P < 0.01, ***P < 0.001. e, Overview STORM reconstruction of assemblies in B/B-stimulated cells expressing hNINJ1–GFP including filamentous structures. Two filaments are highlighted with magenta arrowheads. The indicated regions are magnified on the right. Scale bars, 1 µm. Data in ac,e are representative of at least three independent experiments.

Source data

Atomic structure of the NINJ1 filament

To characterize polymeric NINJ1 at the atomic level, we expressed full-length, wild-type hNINJ1 (residues 1–152) recombinantly in Escherichia coli and extracted it from the bacterial membrane with the detergent n-dodecyl-β-D-maltopyranoside (DDM). The purified protein was visualized by negative staining and cryo-electron microscopy (cryo-EM), showing long filaments with different degrees of bending that occasionally closed to form rings (Fig. 3a and Extended Data Fig. 5a). Quantification of the mass of purified hNINJ1 filaments indicated an average molecular mass of around 1.3 MDa (Extended Data Fig. 5b). The structure of hNINJ1 filaments was determined by cryo-EM. Well-resolved 2D classes representing a top view and initial 3D volumes were used to determine helical parameters via a power spectrum analysis and subsequent helical refinement yielded a final cryo-EM map (Extended Data Fig. 5c–h). The filaments were linear stacks of subunits, with an interval of 20.95 Å and a slight rotation of –1.05° per subunit. The cryo-EM map had a resolution of 3.8 Å, enabling us to build a molecular model of filamentous hNINJ1 (Fig. 3a,b and Extended Data Table 1). The first 38 residues of hNINJ1 remained disordered, as shown by solution NMR relaxation experiments (Extended Data Fig. 6a), in full agreement with predictions from AlphaFold and previous NMR experiments19,20. Indeed, a truncation experiment showed these residues to be dispensable for hNINJ1 filament formation (Extended Data Fig. 6b). The next 103 residues (residues 39–141) were well represented in the maps and could be modelled unambiguously as four α-helices α1–α4 (Extended Data Fig. 5g). Notably, the experimental density comprised two identical hNINJ1 filaments, which were packed together in an antiparallel arrangement (Extended Data Fig. 5h).

Fig. 3: Cryo-EM structure of NINJ1 filament.
figure 3

a, Cryo-EM micrograph showing filamentous hNINJ1 (white arrows) along representative 2D classes. Scale bar, 25 nm. Micrograph representative of 13,124 micrographs from one dataset. b, Organization of hNINJ1 filaments with helices represented as tubes and each subunit shown in a colour gradient (yellow–green–purple). The main interaction interfaces I, II and III are shown below. c, A single hNINJ1 filament subunit, comprising helices α1–α4, with surface representation outlined in light grey. d, Lipophilicity and charge distribution of the hNINJ1 filament. e, Permeability of hNINJ1 proteoliposomes at different protein:lipid molar ratios. Data are mean + s.d. (n = 3 independent experiments). Statistical analysis by one-way ANOVA. ****P < 0.0001.

Source data

The hNINJ1 filament is organized by stacking of adjacent protomers in a fence-like manner (Fig. 3b). The two antiparallel helices α3 and α4 (residues 79–103 and 114–138, respectively) form the core of the filament. These helices are hydrophobic and form a hairpin of transmembrane helices in the inactive form of the protein (Extended Data Fig. 1a). The two N-terminal helices α1 and α2 (residues 44–55 and 58–74, respectively) are separated by a distinct kink at L56 (Fig. 3c). α2 thus adopts a parallel orientation with respect to α3 and α4, whereas α1 protrudes at a nearly 90° angle from the helical bundle and connects to the adjacent protomer via an extensive polymerization interface. The intermolecular contacts include multiple side chain interactions between helix α1 of one protomer and helices α1, α3 and α4 of the neighbouring protomer, of which a salt bridge between the highly conserved residues K45 and D53 is the most prominent (Fig. 3b, I). The interaction also includes newly formed intramolecular contacts via an extensive hydrophobic patch on the amphipathic helix α2 that matches a complementary side chain array on α3 (Fig. 3b, II). Finally, hydrophobic residues on α3 align with a complementary set of residues on α4 of the neighbouring protomer, presumably including a cation–π interaction between K65 and F135 (Fig. 3b, III).

Notably, the experimental structure of NINJ1 in the filament overlaps nearly perfectly with the AlphaFold model for helices α2, α3 and α4, but not for helix α1 (Extended Data Fig. 6c). In the AlphaFold model, α1 and α2 combine to form a single, straight helix. We used molecular dynamics simulations of hypothetical filaments, in which we replaced the individual monomers in the experimentally determined structure with the AlphaFold model. In several subunits, the single α1–α2 helix began to kink at residue 56 and helix α1 restructured itself towards the cryo-EM structure (Extended Data Fig. 6d). Furthermore, we analysed co-evolution of NINJ1, which also underlies the model building in AlphaFold, and focused on the 100 most significant co-evolution pairs. A large majority of these residues pairs corresponds to intramolecular short-range contacts within the NINJ1 monomer. Nine of the residue pairs, however, are in closer spatial proximity between subunits in the filament structure than within one monomer (Extended Data Fig. 6e). In particular, this includes the residue pair F127–G95, which has the highest significance score of all pairs in NINJ1 and which intermolecularly connects helices α3 and α4. These evolutionary data lend further support for the relevance of filamentous NINJ1 in a biological context.

NINJ1 filaments rupture membranes

When analysing the surface hydrophobicity of the NINJ1 filament, we observed that one face of the filament is hydrophilic whereas the other face is hydrophobic (Fig. 3d). These properties directly explain how two hNINJ1 filaments have stacked via their hydrophobic faces to result in the soluble double filament resolved by cryo-EM (Extended Data Fig. 6f). The topology of the filament with one hydrophobic and one hydrophilic face is typical for pore-forming proteins such as gasdermins, perforin or bacterial toxins13,21,22,23,24, and thus readily connects to a functional role in the cell, where these filaments probably cap membrane edges, enabling the rupture of a lipid bilayer membrane. Consistent with this notion, the curved hydrophobic interface measures around 26 Å in height, matching the average thickness of the eukaryotic plasma membrane25. We tested the ability of NINJ1 to increase membrane permeability in liposomes. We reconstituted NINJ1 into proteoliposomes comprising 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphoethanolamine (POPE) and 1-palmitoyl-2-oleoyl-sn-glycero-3-phospho-(1′-rac-glycerol) (POPG) lipids, together with traces of the fluorescent lipid nitrobenzoxadiazole-phosphatodylcholine (NBD-PC). Addition of dithionite, a membrane-impermeable molecule, quenches the fluorescence of NBD-PC only on the leaflets accessible to the bulk solution. Consistent with the hypothesis that NINJ1 induces membrane rupture, the relative permeability of NINJ1 proteoliposomes increased with the protein:lipid ratio of hNINJ1 in the membrane (Fig. 3e and Extended Data Fig. 6g).

Next, we used molecular dynamics simulations to explore the stability of ultrastructural hNINJ1 organizations at membrane edges in silico. Two linear filaments on the opposite edges of a membrane patch remained compact and stable for at least 20 µs in coarse-grained simulations, and for at least 1 µs in all-atom simulations—this was confirmed in two independent replicates each and in control simulations with the experimental double filament (Extended Data Fig. 7a–e). Next, we created a small NINJ1 pore in a membrane in silico, by rearranging a 45-mer filament into a ring. In four coarse-grained simulations of 150 µs length, the NINJ1 polymer remained structurally stable, while showing some variability in the relative orientation of the neighbouring protomers (Extended Data Fig. 7f–h and Supplementary Video 6). In stark contrast, analogous simulations of a ring of truncated NINJ1 lacking helices α1 and α2 underwent a structural collapse and closure of the ring polymers (Extended Data Fig. 7i and Supplementary Video 7). Within a few microseconds, interactions formed between helices α3 and α4 of nearby protomers, and these propagated within tens of microseconds throughout the whole polymer, almost completely closing the ring. Since under similar simulation conditions and duration, structurally defective gasdermin A3 pores undergo substantial reshaping26,27, the observation of stable NINJ1 pores and the collapse of a truncated variant strongly suggest that NINJ1 filaments can cap membrane edges in variable arrangements. Helix α1 thus both stabilizes NINJ1 filaments and confers local plasticity, which could be crucial when perforating densely packed cellular membranes.

Functional validation in cells

To validate our structural model, we designed 14 single-amino-acid mutants and tested their effects on hNINJ1 filament formation in vitro and on cell lysis upon overexpression of mouse NINJ1 (mNINJ1) in cells (Fig. 4a–c). Eight mutants at intermolecular interfaces (K44Q, K45Q, A47L, D53A, G95L, T123L, I134F and A138L) and two mutants at intramolecular interfaces (I84F and Q91A) were designed to potentially break the filament structure, and four mutants at the hydrophobic interface to the membrane (V82F, V82W, L121F and L121W) were designed to be compatible with NINJ1 polymerization. Human and mouse NINJ1 are 98% identical in the structured region of residues 44–138, which includes all 12 mutated sites (Extended Data Fig. 8a). Eight of the ten mutants designed to disrupt filament formation (K45Q, D53A, I84F, Q91A, G95L, T123L, I134F and A138L) reduced or completely abrogated cell lysis upon overexpression in HEK 293T cells, and consistently showed reduced filament formation in vitro (Fig. 4c and Extended Data Fig. 8b,c). Among these, the previously reported K45Q mutant8 showed significantly decreased membrane permeability compared with the wild type in the liposome permeability assay (Fig. 4d and Extended Data Fig. 8d). Mutant A47L was non-functional in HEK 293T cells, but still formed filaments in vitro and increased permeability of proteoliposomes, suggesting that it is involved in additional interactions in cells. Mutant K44Q showed a highly variable phenotype, preventing interpretation. The four mutations in the hydrophobic interface maintaining its hydrophobicity (V82W, V82F, L121F and L121W) did not affect filament formation in vitro as expected (Fig. 4c). Upon overexpression in cells, V82W and L121F were fully functional, whereas mutations V82F and L121W induced lower levels of cell lysis than wild-type NINJ1. This outcome showed that the two sites can tolerate some, but not necessarily all mutations that maintain the hydrophobicity, in full agreement with the structural model. Furthermore, to test the effect of mutating the interaction interfaces between NINJ1 protomers on inflammasome-induced pyroptosis in a physiological setting, we transduced primary Ninj1−/− mouse BMDMs with retroviral vectors expressing either wild-type mNINJ1 or mutants designed to disrupt filament formation and treated the cells with nigericin to activate the NLRP3 inflammasome (Fig. 4e and Extended Data Fig. 8e). In line with the in vitro and HEK 293T overexpression studies, we found that NINJ1 mutants abrogated nigericin-induced release of LDH in reconstituted BMDMs. The sole exception was the A47L mutant, which was non-functional in HEK 293T cells but functional in BMDMs, potentially owing to species-related differences. Overall, the results of in vitro and cell-based mutagenesis were in full agreement with the cryo-EM structure and the expected functional arrangement of NINJ1 filaments at membrane edges.

Fig. 4: The mechanism of NINJ1-mediated PMR.
figure 4

a, Three subunits of filamentous hNINJ1 with overview of residues selected for mutagenesis study (intermolecular interactions, magenta; intramolecular interaction, purple; membrane interactions, green). b, Schematic representation of the residues selected for mutagenesis. c, Cytotoxicity upon overexpression of wild-type (WT) or mutant mNINJ1 in HEK 293T cells. d, Permeability of proteoliposomes containing wild-type and mutant mNINJ1 compared with protein-free liposomes (empty). Data are mean + s.d. (n = 3). e, Release of LDH in primary Ninj1–/– BMDMs reconstituted with wild-type mNINJ1 or different mNINJ1 mutants upon nigericin treatment (1.5 h). Reconstitution with the empty vector and non-transduced Ninj1–/– BMDMs (–) were used as controls. f, Cytotoxicity upon B/B treatment in HeLa cells co-expressing DmrB–caspase-4 and wild-type or mutant mNINJ1. Killing score corresponds to the cytotoxicity, measured by LDH release, normalized against wild-type mNINJ1 control (c) or mock-treated controls (f). Statistical analysis in cf by individual comparison to the control condition highlighted in bold. Data are mean + s.d. and data are pooled from two independent experiments performed in triplicate (c; for K45Q, K45Q, A47L, D53A, V82W, L121W, T123L and A138L mutants), three independent experiments performed in triplicate (c; for mock, WT and V82F, I84F, Q91A, G95L and A138L mutants), and representative of two independent experiments performed in triplicate (c; for L121F mutant), pooled from two independent experiments performed in triplicate (f) or representative of two independent experiments performed in triplicate (e). In c,e,f, multiple plates were used to test all mutants, thus control conditions were included in each of the plates. *P < 0.05, **P < 0.01, ***P < 0.001, **** P < 0.0001; NS, not significant. P values by one-way ANOVA with Dunnet’s multiple comparisons tests. g, Structural model of NINJ1-mediated membrane rupture. Non-activated NINJ1 is randomly distributed in the plasma membrane (PM). Upon activation, NINJ1 polymers lyse the membrane, resulting in the release of cytosolic content (red).

Source data

Finally, we aimed to assess potential dominant-negative effects of the mutants on endogenous NINJ1. We overexpressed mNINJ1 mutants to an intermediate level insufficient to initiate spontaneous cell lysis in GSDMDtg HeLa cells and induced inflammasome activation via DmrB–caspase-4 as before (Fig. 4f and Extended Data Fig. 8f–i). The K45Q, D53A and Q91A mutants had a dominant-negative effect on inflammasome-mediated PMR, without impairing GSDMD activation, which could result from capping or collapsing of endogenous hNINJ1 filaments. Notably, this assay also revealed that the L121F, L121W and T123L mutants, which were dysfunctional in vitro and in overexpression experiments, conferred additional PMR activity compared with controls in the presence of endogenous wild-type protein. Presumably, these mutations weaken the intermolecular interaction in a way that can be tolerated by some of the subunits in a mixed wild-type and mutant filament, but not in every repeating unit of a filament comprising only mutant protein. In summary, mutagenesis confirmed that the formation of NINJ1 filaments in vitro correlates with the ability to induce PMR in both human and mouse cells.

Discussion

Here we combined insights from super-resolution microscopy, cryo-EM, mutation analysis and molecular dynamics simulation to provide an atomic model for membrane rupture by NINJ1 (Fig. 4g). In live cells, NINJ1 is monomeric in the cellular membrane, with helices α1 and α2 on the extracellular side and the pair of α3 and α4 integral to the membrane (Extended Data Fig. 1a). During cell death, amphipathic helices α1 and α2 insert into the membrane to adopt the kinked conformation, bridging neighbouring protomers to form larger polymers. The resulting higher-polymeric assemblies promote membrane rupture by capping membrane edges, thus stabilizing membrane lesions of variable size and morphology through which LDH, large danger-associated molecular patterns (DAMPs) and other cellular content are released into the extracellular milieu. Although single filaments might be sufficient to damage membranes, it is also plausible that NINJ1 forms double filaments that open up in a zipper-like manner in response to osmotic pressure to form membrane lesions.

The trigger that causes the transition of NINJ1 from the inactive state to the active state remains unknown. The proposed polymerization mechanism raises the interesting possibility that the membrane composition might contribute at least partially an activation signal. During cell death, negatively charged phosphatidylserine becomes exposed on the cell surface28,29, which might be recognized by helices α1 and α2 of NINJ1. Indeed, lipid binding experiments and dye release assays show that a peptide corresponding to helices α1 and α2 interacts specifically with POPS-containing membranes, and molecular dynamics simulations show the same effect (Extended Data Fig. 9a–c). Membrane-composition sensing as a potential activation mechanism of NINJ1 is thus a promising avenue for future work.

In summary, active NINJ1 has a unique structure with long, α-helical filaments capping membrane edges. Whereas the β-sheet structure of the GSDMD pore has a limited pore size that allows interleukin release while retaining larger molecules13, the membrane openings or lesions caused by NINJ1 filaments are essentially unconstrained in size. NINJ1 lesions appear functionally related to the large superstructures of activated mitochondrial Bax and Bak, in that they both lyse membranes30. Although atomic structures of Bax or Bak in a pore-forming conformation are not currently available, we speculate that the helical hairpin of α3 and α4 in the NINJ1 filament might show functional and structural resemblance to the helical hairpin of α5 and α6 in activated Bax. NINJ1 was initially reported as an adhesion molecule, induced after sciatic nerve injury and promoting axonal growth31,32,33. Given that NINJ1 drives cell death and the release of DAMPs, and close links exist between cell death, inflammation and tissue repair34, it is conceivable that NINJ1 has an indirect role in promoting axonal growth either by causing inflammation or the release of stimulatory molecules. Conversely, it is also possible that NINJ1 has a dual role, serving in both cell–cell adhesion and as a breaking point for membranes at strong osmotic pressure35. The structure, along with the mutagenesis studies provide possible explanations for why NINJ2 is not able to functionally replace NINJ1, despite a high degree of homology8. The two proteins differ by a few amino acids in the structured part, and by a large deletion and multiple mutations in the unstructured N terminus (Extended Data Fig. 9d). Among the differences in the structured region is residue 47, which is an alanine in NINJ1 and a valine in NINJ2. The A47L mutation made NINJ1 dysfunctional in HEK cells (Fig. 4c), providing a possible explanation for the dysfunction of NINJ2, and it is likely that several of the other differences lead to additional perturbations in function. G93, L57 and Q63 in NINJ1 correspond to V, F and R in NINJ2, respectively, each of which would probably cause steric clashes that might prevent filament formation. Of note, NINJ1 represents yet another occurrence of filament formation in pyroptotic pathways, along with the soluble filaments of PYD and CARD domains of ASC, NLRP3 and caspase-130,36,37,38,39. The NINJ1 filament represents an elegant biophysical mechanism of cellular disintegration, and knowledge of its atomic structure opens opportunities for therapeutic interventions in cancer, infection and inflammatory diseases.